Patent Application
20110314031
As the use of network-based publication systems and marketplaces, such as on-line commerce services or auction services expands, and the volume of item listings in such applications increases, the speed, ease and convenience with which information can be retrieved from such marketplaces increases in importance to customers.
Item listings in such network-based marketplaces typically include details of a particular item which is e.g. up for sale or auction and these details are typically stored in text format and include a description of the item together with other information, such as the price, useful to a potential buyer in assessing the item listing. In addition, item listings often include visual material related to the item, typically in the form of a photograph, drawings or video clips.
The use of images in this context has customarily been limited to the provision of information about item listings to customers, but it would be useful to employ images associated with item listings for additional purposes, such as for image-based searching or for the automatic identification of images for fraud protection purposes.
The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary foe. Some embodiments are illustrated by way of example and not limitation in the figures of the accompanying drawings in which:
In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of some example embodiments. It will be evident, however, to one skilled in the art that the present embodiments may be practiced without these specific details.
The index sets are associated with their respective images in a so-called reverse indexing arrangement, in which the index database contains a single entry for each unique index set or feature vector, with all the images that contain that particular index set or feature vector being listed against the entry.
When, for instance, a query image is subsequently presented in order to search for identical or similar images in the image database, the query image is itself parsed or processed to generate multiple feature vectors or sets of index values for the query image. The index sets for the query image are thus generated in a similar manner to the generation of index sets for the index database.
To identify images similar to the query image, all of the query image's index sets are compared to index sets in the index database. When commonality is identified between a query image index set and a database index vector, a hit count is incremented for each of the images associated with that index set in the index database. The database images are ranked in order of hit count, so that the result of the search are those item listings whose images have index sets showing the greatest commonality with the index sets of the query image.
In an example embodiment, the index sets are generated by first performing an edge detection algorithm on the image and then normalizing the image. Thereafter, the image is divided into a raster of cells or sub-portions at a resolution which is usually significantly lower than its native pixel resolution, so that each cell or sub-portion comprises many image pixels. The raster may typically be a 10 by 10 grid of cells. Next, an index value is assigned to each cell based on an image attribute, typically a light intensity value, of the underlying image pixels. The result is therefore a grid of index values. Thereafter, sets of index values feature vectors) are created from respective overlapping portions of the grid of index values. This process can be explained as a sliding window that is moved over the grid to isolate two dimensional selections or portions of index values at a time, the portions overlapping and covering the entire grid. The index values in each selection or portion together form one feature vector or set of index values. In an example embodiment, the moving window is three by three cells in size, so that each feature vector is constituted by nine index values.
The system and method thus provides for image comparison in a publication system, such as a network-based marketplace, which requires significantly less processing power, and is therefore faster than image comparison based on native image data, while returning results with relatively high accuracy. The method is furthermore relatively robust and insensitive to scaling and resolution loss.
One example embodiment of a distributed network implementing image-based indexing for item listings in a network-based marketplace is illustrated in the network diagram of
Turning specifically to the network-based marketplace platform 12, an Application Program Interface (API) server 24 and a web server 26 are coupled to, and provide programmatic and web interfaces respectively to, one or more application servers 28. The application servers 28 host one or more marketplace applications 30 and payment applications 32. The application servers 28 are, in turn, shown to be coupled to one or more databases servers 34 that facilitate access to a number of databases, in particular an item listing database 35, an image database 36, and an index database 37. The item listing database 35 stores data indicative of item listings for items which are offered for sale or auction on the platform 12. Each item listing includes, inter alia, a text description of the relevant item and metadata categorizing the item. The image database 36 includes images associated with respective item listings in the item listing database 35. The images in the image database 36 may be standard format image files such as JPEG files. The index database 37 contains index data relating to images in the image database to permit image-based searching of the image database 36. The format of index data in the index database is described in more detail below.
The marketplace applications 30 provide a number of marketplace functions and services to users that access the marketplace platform 12. The payment applications 32 likewise provide a number of payment services and functions to users. The payment applications 32 may allow users to quantify for, and accumulate, value (e.g., in a commercial currency, such as the U.S. dollar, or a proprietary currency, such as “points”) in accounts, and then later to redeem the accumulated value for products (e.g., goods or services) that are made available via the marketplace applications 30. While the marketplace and payment applications 30 and 32 are shown in
Further, while the system 10 shown in
The web client 16, it will be appreciated, accesses the various marketplace and payment applications 30 and 32 via the web interface supported by the web server 26. Similarly, the programmatic client 18 accesses the various services and functions provided by the marketplace and payment applications 30 and 32 via the programmatic interface provided by the API server 24. The programmatic client 18 may, for example, be a seller application (e.g., the TurboLister application developed by eBay Inc., of San Jose, Calif.) to enable sellers to author and manage listings on the marketplace platform 12 in an off-line manner, and to perform batch-mode communications between the programmatic client 18 and the network-based marketplace platform 12.
A number of fixed-price applications 46 support fixed-price listing formats (e.g., the traditional classified advertisement-type listing or a catalogue listing) and buyout-type listings. Specifically, buyout-type listings (e.g., including the Buy-It-Now (BIN) technology developed by eBay Inc., of San Jose, Calif.) may be offered in conjunction with an auction-format listing, and allow a buyer to purchase goods or services, which are also being offered for sale via an auction, for a fixed-price that is typically higher than the starting price of the auction.
Store applications 48 allow sellers to group their listings within a “virtual” store, which may be branded and otherwise personalized by and for the sellers. Such a virtual store may also offer promotions, incentives and features that are specific and personalized to a relevant seller.
Reputation applications 50 allow parties that transact utilizing the network-based marketplace platform 12 to establish, build and maintain reputations, which may be made available and published to potential trading partners. Consider that where, for example, the network-based marketplace platform 12 supports person-to-person trading, users may have no history or other reference information whereby the trustworthiness and credibility of potential trading partners may be assessed. The reputation applications 50 allows a user, for example through feedback provided by other transaction partners, to establish a reputation within the network-based marketplace platform 12 over time. Other potential trading partners may then reference such a reputation for the purposes of assessing credibility and trustworthiness.
Personalization applications 52 allow users of the marketplace platform 12 to personalize various aspects of their interactions with the marketplace platform 12. For example a user may, utilizing an appropriate personalization application 52, create a personalized reference page at which information regarding transactions to which the user is (or has been) a party may be viewed. Further, a personalization application 52 may enable a user to personalize listings and other aspects of their interactions with the marketplace and other parties.
In one embodiment, the network-based marketplace platform 12 may support a number of marketplaces that are customized, for example, for specific geographic regions. A version of the marketplace may be customized for the United Kingdom, whereas another version of the marketplace may be customized for the United States. Each of these versions may operate as an independent marketplace, or may be customized (or internationalized) presentations of a common underlying marketplaee.
Navigation of the network based-marketplace may be facilitated by one or more navigation applications 56. For example, a keyword search application 57 enables keyword searches of listings published via the marketplace platform 12. Similarly, an image search application 59 enables an image-based search of item listings published via the marketplace platform 12. To perform an image-based search, a user will typically submit a query image, whereupon the image search application 59 may compare the query image to images in the image database to produce a result list of item listings based on a similarity ranking between the query image and the images associated with the respective item listings. The comparison ranking is established by parsing or processing the query image to provide index data, and thereafter comparing the query image's index data to pre-compiled index data for the listing images, as described in more detail below. A browse application allows users to browse various category, catalogue, or inventory data structures according to which listings may be classified within the marketplace platform 12. Various other navigation applications may be provided to supplement the search and browsing applications.
In order to make listings, available via the network-based marketplace, as visually informative and attractive as possible, as well as to enable image-based searching, the marketplace applications 30 may include one or more imaging applications 58, which users may use to upload images for inclusion within listings. Images thus uploaded are stored in the image database 36, each image being associatively linked to at least one item listing in the item listing database 35. One of the imaging applications 58 also operates to incorporate images within viewed listings. The imaging applications 58 may also support one or more promotional features, such as image galleries that are presented to potential buyers. For example, sellers may pay an additional fee to have an image included within a gallery of images for promoted items.
The marketplace platform 12 may also include an image indexing application 61 to parse or process images uploaded via the image application 58, as well as to parse or process query images submitted via the image search application 59. The result of processing images by the image indexing application 61 is index data which is stored in the index database 37. Particular processes for indexing images, as well as the format of index data, are discussed in more detail below.
Listing creation applications 60 allow sellers conveniently to author listings pertaining to goods or services that they wish to transact via the marketplace platform 12, and listing management applications 62 allow sellers to manage such listings. Specifically, where a particular seller has authored and/or published a large number of listings, the management of such listings may present a challenge. The listing management applications 62 provide a number of features (e.g., auto-relisting, inventory level monitors, etc.) to assist the seller in managing such listings, One or more post-listing management applications 64 also assists sellers with a number of activities that typically occur post-listing. For example, upon completion of an auction facilitated by one or more auction applications 44, a seller may wish to leave feedback regarding a particular buyer. To this end, a post-listing management application 64 may provide an interface to one or more reputation applications 50, so as to allow the seller conveniently to provide feedback regarding multiple buyers to the reputation applications 50.
Dispute resolution applications 66 provide mechanisms whereby disputes arising between transacting parties may be resolved. For example, the dispute resolution applications 66 may provide guided procedures whereby the parties are guided through a number of steps in an attempt to settle a dispute. In the event that the dispute cannot be settled via the guided procedures, the dispute may be escalated to a third party mediator or arbitrator.
A number of fraud prevention applications 68 implement various fraud detection and prevention mechanisms to reduce the occurrence of fraud within the marketplace. One of the fraud prevention applications 68 may include automatic image comparison, by use of index data produced by the image indexing application 61 and stored in the index database 37. Such image comparison may be used by the fraud prevention application 68 automatically to detect listing images similar to the query image, and to alert a fraud assessor to such image listings, so that the human assessor can examine the identified item listing for assessing whether or not the identified item listing is a fraudulent listing.
Messaging applications 70 are responsible for the generation and delivery of messages to users of the network-based marketplace platform 12, such messages for example advising users regarding the status of listings at the marketplace (e.g., providing “outbid” notices to bidders during an auction process or providing promotional and merchandising information to users).
Merchandising applications 72 support various merchandising functions that are made available to sellers to enable sellers to increase sales via the marketplace platform 12. The merchandising applications 72 also operate the various merchandising features that may be invoked by sellers, and may monitor and track the success of merchandising strategies employed by sellers.
The network-based marketplace platform 12 itself, or one or more parties that transact via the marketplace, may operate loyalty programs that are supported by one or more loyalty/promotions applications 74. For example, a buyer may earn loyalty or promotions points for each transaction established and/or concluded with a particular seller, and be offered a reward for which accumulated loyalty points can be redeemed.
The receiving module 80 is operatively connected to a processing module 86, which may form part of the image indexing application 61, for processing images to generate index data for the images. An example method of processing images by the processing module 86 to generate the index data is described in more detail below. An image listing module 82, which may form part of the listing creation application 60 of
The processing module 86 is configured to parse or process images submitted to it to generate index data in the form of a number of a feature vector or index sets 88 for each image. In an example embodiment, each index set 88 comprises a sequence of nine index values, as shown in
The processing module 86 is in communication with the index database 37, to communicate index data generated by the processing module 86 to the index database 37 for storage, In an example embodiment, index data is related in the index database 37 to images stored in the listing image database 36 in a reverse-indexing format. The data format in the index database 37 may be in a spatial data structure, such as a two-dimensional k-d tree, to facilitate searching of the database 37. As illustrated in
Each entry in the index database 37 is linked to at least one image in the listing image database 36. Further, the images stored in the image database 36 are linked to associated item listings in the item listing database 35. As discussed above, each item listing may comprise information about a particular item which is offered for sale or auction on the marketplace platform 12, such as a description of the item, at least one category of the item, a brand name of the item, etc. In the example embodiment, entries in the respective databases 35 to 37 are linked by use of image filenames as linking data. A person skilled in the art will appreciate that, in other embodiments, any appropriate data structure, e.g. relational databases or tables, may be used to link images in the image database 36 to, on the one hand, respective item listings and to, on the other hand, index data in the index database 37.
A comparison module 90 is in communication with both the processing module 86 and the index database 37, to compare index data for a query image with the index data in the index database 37 for finding database images similar to the query image. The comparison module, in use, produces a ranking of database images according to similarity to the query image. This ranking is achieved by comparing the feature vectors 88 of the query image, which is generated by the processing module 86, with feature vector entries in the index database 37 in accordance with a comparison method which is described in more detail below.
The system 10 also includes a result module 92 for delivering to a user the results of a search query which includes a query image. The result module 92 is configured to return query search results as item listings ranked in order of the similarity of their respective images to the query image, as determined by the comparison module 90. The search results may be delivered by the user in a format of the user's choosing, such as, for instance, via email or in http format on a web browser application.
The image may be an electronic image in standard file format, such as JPEG, which comprises a raster of pixels. Each pixel may have hue, saturation, and intensity values, in conventional fashion. It will be appreciated that images which were submitted earlier and which are stored in the image database 36, but which have not been indexed, may be provided to the processing module 86 by the database crawler 84 of
The submitted image is then processed, at block 122, to generate index data for enabling index searching of database images. Processing of the image will be described at the hand of an example image 200 illustrated in
The normalised edge image 202 is then partitioned or divided, at block 108, into cells or sub-portions 207, to form a grid 204 (
Thereafter, an index value 208 is assigned, at block 110, to each cell 207 based on an image attribute of the underlying image pixels of the cell 207. In this example, the image attribute is intensity, typically measured on a scale of 0 to 255, or alternatively 0 to 100, where a pixel having a zero intensity is white and a pixel having intensity value of 255 (or 100, as the case may be) is black. The index value 208 assigned to each cell 207 may thus be the average tight intensity value of the pixels constituting each cell 207. The output of operation 110 is therefore a ten-by-ten grid 206 of index values 208 based on the intensity values of the respective cells 207 (
In other example embodiments, other image attributes can be used as well as or instead of the intensity value. For instance, colour values of the cells 207 may be calculated and indexed together with or instead of the intensity index values. An index value may for instance assigned to each cell 207 based on the average hue of the cell's pixels. Instead, separate grids may be produced for red, green and blue colour spaces, and index values based on the intensity values of the respective colours in the cells may be assigned to the cells.
At block 140, feature vectors or index sets 88 are compiled from the grid 206. Compilation of the index sets 88 comprises iteratively isolating portions 210 of the grid 206 and listing the index values 208 in each portion 210 in sequence, to provide an index set 88. Compilation of the index sets 88 can thus be described as a sliding, overlapping mask or window 210 which is three-by-three index values in size, and which iteratively isolates all possible contiguous three-by-three selections in the grid 206, to generate respective index sets 88. Each index set 88 thus comprises a sequence or vector of nine index values 208. Although only two of these index sets 88 are shown in
The index sets 88 thus generated are incorporated in the index database 37. As explained above with reference to HG. 3, the index database 37 comprises a single entry for each unique feature vector or index set 88, with all images which contain that index set 88 being listed in the entry. To this end, each index set 88 generated in operation 112 is processed by first establishing, at block 114, whether or not the index database 37 already includes an entry for that particular index set 88. If the determination at block 114 is in the affirmative, the image which is the subject of current processing is linked, at block 118, to the existing entry in the index database 37 by including the image filename in the listing of images in the respective database entry, if, however, the determination is in the negative and there is not yet a database entry for the index set 88 under consideration, a new index set entry is created, at block 116, in the index database 37. It will be appreciated that such a new database entry will have only the current image associated with the particular index set 88. This database writing sequence is looped, at block 120, through all of the index sets 88 generated at operation 112, therefore being performed 64 times in the example embodiment having a ten-by-ten grid 206.
The query image is processed, at block 122, by the processing module 86 to generate index sets 88 for the query image in a manner identical to indexing of images in accordance with the method 100 of
After generation of the query index sets 88 from the query image, the query feature vectors or index sets 88 are compared to the index data in the index database 37 to indentify images similar or identical to the query image. To this end, the comparison module 90 processes each of the query index sets 88 in turn. The comparison module 90 steps or loops, at block 146, through the index values 208 of a particular index set 88 to find, at block 142, all index set entries in the index database 37 that share that index value 208. For each index entry identified as sharing the particular index value 208 under consideration, a hit count is incremented for each of the images associated in the index database 37 with the identified database entry. The database images are then ranked in descending order according to hit count.
In an example where the first query index set 88 is [24, 12, 13, . . . , 4], the first index value 208 is 24. If the index database 37 includes the following entries:
1. [0, 0, 24, 16, 26, . . . ]=IMG#3224, IMG#6739
2. [36, 48, 18, . . . ]=IMG#644, IMG#2542
3. [24, 12, 0, 0, . . . ]=IMG#3224, IMG#2143,
the first iteration of operation 142, in respect of value 24, will result in an increment in the hit count of the images in data entries 1 and 3 above. The second iteration of operation 142, which will be in respect of value 12, will result in the incrementing, at block 144, of the images in data entry 3 above. The results of looping through all the index values 208 of query index set 88 shown above will be IMG#3224=3 hits; IMG#2143=2 hits, with the remainder of the listed images registering a single hit, apart from IMG#644 and IMG#2542 which will have no hits registered against them.
After repeating operations 142 to 146 for all of the index values 208 of one of the index sets 88, the process loops, at block 148, to the next query index set 88. Operations 142 to 148 are thus repeated until all of the index values of all of the query index sets 88 have been compared to the index database entries, the hit counts being aggregated to provide a ranking of images by hit count, at block 150.
In other embodiments, the comparison of index sets may include comparing all of the index values of query index sets with all of the values in the respective pre-compiled index sets forming database entries in the index database 37, to that a hit is registered only if there is complete overlap between the index values of, on the one hand, the query index set, and, on the other hand, the index values of the particular database entry. In yet further embodiments, the comparison of feature vectors or index sets may include matching not only the values of the query index sets to database entry index sets, but also matching the sequence of index values in the respective index sets. A hit will thus be registered only if the query index set matches a database entry's feature vector or index set 88 exactly, in other words if both the values and the sequence of the respective index sets are identical. To promote processing speed and efficiency when performing exact feature vector or index set matching, the index value range may be reduced in scale, so that the index values, for example, range in value from 0 to 10 instead of, for example, from 0 to 100.
Instead, or in addition, the comparison operation may include a weighting of the hit count based on the position of the respective index sets in the image. In other words, hits may be assigned weights based on adjacency of the index sets in the image. Two matching index sets which were compiled from image portions or windows which are in adjacent or identical locations in the grid may therefore result in a higher weighted hit, while a lower weighted hit may be registered if the respective image portions or windows are less adjacent.
As mentioned above, pre-complied feature vectors or index sets 88 may be stored in the index database in a data structure like a k-dimensional tree, also known as a k-d tree, Comparison of a query index set or feature vector may in such cases comprise performing a nearest neighbour search in the k-dimensional tree.
The hit counts of the images are passed by the comparison module 90 to the result module 92. The result module 92 then displays to the user, at block 152, the results of the search. The search results are provided as a list of item listings extracted from the item listing database 35, the displayed item listing being the item listings linked to the top ranked images, as identified by the comparison module 90.
The system 10 and methods 100, 130 described above provides for effective image-based searching in the network-based marketplace. Indexing of the images in the image database 36 in accordance with the described example embodiment permits similarity comparison of the query image with large numbers of database images without requiring prohibitive processing power or time.
In addition to use of the indexing method 100 for user-initiated image-based searching of the database 36, it may, in other embodiments, be used for fraud prevention applications in the network-based marketplace. In such embodiments, the fraud prevention application 68, shown in
In other embodiments, the processing operation 122 to generate index sets 88 may differ in a number of aspects, some of which are described with reference to
In another embodiment, such image variation may be performed upon searching instead of, or in addition to, image variation during indexing, In such case, a query image may thus be processed to produce multiple image variations, index sets 88 thereafter being generated for each of the variations and being compared to the index database 37.
In the embodiment illustrated in
In another embodiment, illustrated in
In yet another embodiment, illustrated in
In another embodiment, the partitioning resolution for indexing may be determined by a category of the relevant item listing. For example, the method may include categorising an item listing upon creation, determining the item listing's category before processing the image, at block 108, and selecting the partitioning resolution based on the item listing category. For instance, apparel may be partitioned at a 10×10 resolution, while electronic devices may be partitioned at a 15×15 resolution.
It will further be appreciated that the system and methodology described above can be applied to video content as well as or instead of image data. The method may in such case include extracting images from video files, and processing the extracted images in accordance with the example embodiments described herein. Image extraction from such video content may include automatically identifying scene changes in the video content, e.g. by comparison of successive frames or images in the video content, and extracting images or frames at the start of such scene changes.
The example embodiments described herein address some of the technical challenges associated with effective processing of images linked to item listings. For instance, image databases of network-based marketplaces are often very large, comprising millions of images, so that the time and/or processing power consumed by conducting a search or image comparison in the database is prohibitive. However, a comparison of index data generated for the query image with index data of database images is considerably less resource intensive.
In another example embodiment, there is provided a system and a method to index images associated with item listings in a network-based marketplace, so that subsequent similarity search or comparison operations are performed on index data instead of on base images stored in an image database. Images in the image database are thus indexed by parsing or processing the images for feature extraction. The feature extraction may include generating multiple sets of index values associated with each image. Each set of index values is also referred to as a feature vector or an image fingerprint.
Digests for various image features can be pre-computed and stored in the image database for subsequent recall, as a recall set, for comparison to a query for an item, for example, at run time. As examples, a digest can include edge information as alluded to above, color information, pattern, quality, texture, and the like. Digests can also include textual information as well as image information such as item attributes that are either determined by the system at the time of listing, or are specified by the seller. However, here “textual” does not mean merely words typed by the user. “Textual” means text information to be stored about the item the seller is listing, including item title, and also attributes of the item. This can be considered “structured” and “unstructured” item data. For example, unstructured data would include item title and item description, which could include whatever data the seller wants to include. Structured data would be data such as the item category (e.g., shoes, handbags, and the like) and item aspects such as brand, material, and the like, where each piece of information can be stored as separate name-value pairs. So, if the listing item is a Coach handbag with title “New brown Coach purse”, the digest may include the following terms as textual elements of a pre-computed digest (new, brown, Coach, purse, Brand:Coach, Material:leather, Style:satchel).
As an example of an edge part of a digest, the image can be converted into a grid as discussed in detail below, where each box in the grid contains a value indicating the presence of a signal, such as a pixel, which, when integrated across all grids of interest, indicate a strong edge in that region of the image. The edge digest is stored as a string of buckets representing each element in a matrix. This is discussed in more detail below.
In performing a color process, it is desirable in pre-computing digests, or for building a recall set for comparison with a query, to determine whether a first pixel in a number of pixels bears a positive likeness to a second pixel or a negative likeness to the second pixel. This can be seen in
Returning to
Edges of the image 111 can be found by a technique called water shedding. In a recomposing process, each area of an image is filled with color. This can be visualized generally as if the image were a terrain map and the color were water. The points of the watershed image 113 of
Next, a segment graph 115 is created from the watershed image 113. Small segments are joined to the most-similar neighbor. Very similar adjacent segments are also joined. Then the border segments are marked as part of the background unless the average pixel is in the middle third or most of the segment's pixels are in the middle third. The background is zeroed out from the original image, and a hue histogram is calculated to get a compressed color representation 117 for the image.
As discussed previously, the color representation 117 of the image may be accomplished with hue level color bucketing using an EELS cone. An HLS cone is seen in
The bottom region 125 of the cone is black.
At the top level of the cone, the center region 127 of the circle 121 is white.
The middle region 129 of the cone is grey.
On the rest of the image, the outer edge of the circumference 123 is divided into 16 colour buckets.
For each image, such as the image 111 of
If the histograms are looked upon as vectors, the standard Euclidian distance between the two vectors could be calculated. Unfortunately, this yields an unreal effect because Euclidian distance indicates that each color bucket is weighed the same. That is, red is just as similar to orange in Euclidian distance as red is to purple, or as red is to green, or as red is to black. But human perception does not operate in that manner; color lies on a spectrum (red, orange, yellow, green, blue, indigo, and violet) in human vision such that red is closer to orange and yellow than red is to green. Therefore the color buckets should be weighted by the distance from each other as well. So the color buckets can be organized linearly around the circumference 123 of hue circle 121, as seen in
The color similarity assumes a fixed distance between black, grey, and white (four units) and a distance between each color bucket is a unit distance in the circular scale at the circumference 123. Stated another way, the circular hue space shown in
This provides an improvement when two histograms are compared to determine how far these histograms are from each other; that is, it enables the system to determine how much visual difference obtains between two different histograms. In the comparison of histogram vectors, the color buckets are weighted by the distance between them by providing, in the similarity comparison, that if a given color bucket is next to another given color bucket, as in
Texture detection can be performed by the system using grey level co-occurrence analysis. The system can convert an image to a grey scale, for example, a 16-color grey matrix, numbered one through sixteen on each coordinate, as illustrated in
The above matrix (for the pre-computed digest, and also for the query image at run time) can be transformed to emphasize high contrast data, by the following equation:
n
x,y
=m
x,y((|x−y|)+1)2
mx,y is the value in the original matrix at position (x,y); and
x and y represent grey values.
So, referring again to
As mentioned previously, when comparing two images, such as a query input from the user interface and digests from index database 37, the two images are compared by taking edge information, then placing one image over the other, and then determining the differences in the images, as discussed above. However, the process is not robust for rotation. For example, and with reference to
This rotation can be accomplished by using principal component analysis (PGA). PCA is a mathematical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of uncorrelated variables called principal components. The number of principal components is less than or equal to the number of original variables. This transformation is defined in such a way that the first principal component has as high a variance as possible (that is, accounts for as much of the variability in the data as possible), and each succeeding component in turn has the highest variance possible under the constraint that it be orthogonal to (uncorrelated with) the preceding components. Principal components are guaranteed to be independent only if the data set is jointly normally distributed. PCA is sensitive to the relative scaling of the original variables. However, in this instance, PCA is used to determine an angle or a direction in which the images are each spread out the most so that similar objects would have similar alignments. This method can be performed by finding the two major Eigen vectors of each edge detected of images 145, 147, one the listing at digest creation time, and the other the query image at runtime. Then the system aligns one of the major Eigen vectors, for example the first major Eigen vector, K1 in
In addition, image features for shape may not be translation invariant. For example, a similarity measure for an image 140 in
At digest create time, the system can operate on a large corpus of listing images linked to sub-images to generate vectors as described, providing a large pool of sub-images, perhaps in the millions. The system then runs a k-means clustering algorithm on the corpus of sub-images. K-means clustering is a method of cluster analysis. A k-means cluster algorithm assigns each point to the cluster whose center (also called a centroid) is nearest. The center is the average of all the points in the cluster—that is, its coordinates are the arithmetic mean for each dimension separately over all the points in the cluster.
Example: The data set has three dimensions and the cluster has two points:
X=(x1,x2,x3) and Y=(y1,y2,y3).
Then the centroid Z becomes Z=(z1,z2,z3),
Where
Applying a k-means cluster algorithm, to, for example, millions of nine-dimensional vectors during digest creation will reduce the vectors to a reasonable number of centroid points, perhaps as few as one thousand to ten thousand. These centroid points can be viewed as code points in a code book for each of the respective vectors. At runtime, when a user submits an image query to the system, the system looks at each sub-image (such as sub-image 142 in
As discussed above, a listing query sometimes referred to as “query image”) can be subsequently presented by web client 16 of
Textual information which could include, for example, title and attributes, is stored with an item and can be used in the first-pass query based on textual similarity. For example, if the textual information of a pivot item included Title=“New brown Coach purse” and aspects Brand:Coach, Material:leather and Style=satchel, all these values would be passed to a RANK operator as the “text query”. The RANK operator compares the text query against images stored as pre-computed digests in index database 37 using a TF-IDF like algorithm, no each term is weighted according to inverse document frequency, meaning terms that are rare have higher weight. For example, if one in two items in the handbags category contained the word “new”, then “new” would have an IDF score of 2, whereas if Brand=Coach appears only 1 in 20 times in the handbags category, then this term would have a weight of 20. The RANK operator looks for items that match any of these terms, and gives a score according to number of matches and associated term weights. Then the top N items from this step can be re-sorted using a second pass sort using a second ranking method, such as a “Best Match” method like those disclosed in copending application Ser. No. 12/476,046, entitled “Methods and Systems for Deriving a Score with which Item Listings are Ordered.” An example of a second pass ranking that re-sorts these top N items might use a combination of image comparison operations (for example: distance between color histograms, discussed above, cosine similarity of edge matrices, pattern comparison, and the like), textual similarity, by time (sale ending soonest); by newly listed; by price+shipping cost (lowest first); by price+shipping cost (highest first); by best match; by price (highest first); by price (lowest first), and seller or item quality measurements to produce a final ordering which is presented to users. In this case, as an example, the profiles for image similarity search can be:
w
—1*ColorSim(colordigest)+w—2*EdgeSim(edgedigest)+w—3*SimilarityScore
where
w is a weight customizable by pivot item category,
ColorSim is the distance between color digests of the two images
EdgeSim is the distance between edge digests of the two images
SimilarityScore is the score computed above in the first-pass sort
As discussed previously, both color and edge digests are represented as histograms measuring the number of pixels falling into distinct bins: for color the bins are regions of a given color space, and for edge the bins are gradients of a certain magnitude. There are many ways of finding distances between two histograms. One way is to use histogram intersection for color, and cosine similarity for edge.
The weights “w” on each factor are optimized using feedback from users as a way of optimizing sorts by item type or category. Datasets can be used comprising idealized orderings as ranked by human judges, and/or by click-through patterns from an ecommerce website to optimize the importance of each factor to users by product category. Then when a user selects an item for a query, the optimum weights for that item's category can be used to rank the results to be presented to the user.
The method of obtaining weights w, above, by human judges is discussed with respect to
An example operation of the image similarity search function described above can be seen from
Continuing with the discussion of
if desired, the “More Like This” similarity search can be designed so that one can enter the “More Like This” similarity search by way of a similarity search at “All Categories” button 165, for example, of
The system can be designed such that at any level of the above “More Like This” similarity searches, regardless of the method of entering the search, the system can allow the user to switch from one category of item to another category of item. For example, when in the “More Like This” similarity search on
Continuing with the above discussion, a textual “More Like This” search could be also be entered through box 190 of
At this point the system can test for an end signal and if an end signal is detected, the method ends; if no end signal is detected the system continues to decision [point 197 to test for a change of item category. If the system detects a change of item category as at the Yes decision of operation 197, such as the user entering a new category by any of buttons 161, 165, or box 167 of
It will further be appreciated that the system and methodology described above can be applied to video content as well as or instead of image data. The method may in such case include extracting images from video files, and processing the extracted images in accordance with the example embodiments described herein. Image extraction from such video content may include automatically identifying scene changes in the video content, e.g. by comparison of successive frames or images in the video content, and extracting images or frames at the start of such scene changes.
The example embodiments described herein address some of the technical challenges associated with effective processing of images linked to item listings. For instance, image databases of network-based marketplaces are often very large, comprising millions of images, so that the time and/or processing power consumed by conducting a search of image comparison in the database is prohibitive. However, a comparison of index data generated for the query image with index data of database images is considerably less resource intensive.
Certain embodiments are described herein as including logic or a number of components, modules, or mechanisms. A component is a tangible unit capable of performing certain operations and may be configured or arranged in a certain manner. In example embodiments, one or more computer systems (e.g., a standalone, client or server computer system) or one Or more components of a computer system (e.g., a processor or a group of processors) may be configured by software (e.g., an application or application portion) as a component that operates to perform certain operations as described herein.
In various embodiments, a component may be implemented mechanically or electronically. For example, a component may comprise dedicated circuitry or logic that is permanently configured (e.g., as a special-purpose processor) to perform certain operations. A component may also comprise programmable logic or circuitry (e.g., as encompassed within a general-purpose processor or other programmable processor) that is temporarily configured by software to perform certain operations. It will be appreciated that the decision to implement a component mechanically, in dedicated and permanently configured circuitry, or in temporarily configured circuitry (e.g., configured by software) may be driven by cost and time considerations.
Accordingly, the term “component” should be understood to encompass a tangible entity, be that an entity that is physically constructed, permanently configured (e.g., hardwired) or temporarily configured (e.g., programmed) to operate in a certain manner and/or to perform certain operations described herein. Considering embodiments in which components are temporarily configured (e.g., programmed), each of the components need not be configured or instantiated at any one instance in time. For example, where the components comprise a general-purpose processor configured using software, the general-purpose processor may be configured as respective different components at different times. Software may accordingly configure a processor, for example, to constitute a particular component at one instance of time and to constitute a different component at a different instance of time.
Components can provide information to, and receive information from, other components. Accordingly, the described components may be regarded as being communicatively coupled. Where multiple of such components exist contemporaneously, communications may be achieved through signal transmission (e.g., over appropriate circuits and buses) that connect the components. In embodiments in which multiple components are configured or instantiated at different times, communications between such components may be achieved, for example, through the storage and retrieval of information in memory structures to which the multiple components have access. For example, one component may perform an operation, and store the output of that operation in a memory device to which it is communicatively coupled. A further component may then, at a later time, access the memory device to retrieve and process the stored output. Components may also initiate communications with input or output devices, and can operate on a resource (e.g., a collection of information).
The various operations of example methods described herein may be performed, at least partially, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors may constitute processor-implemented modules that operate to perform one or more operations or functions. The modules referred to herein may, in some example embodiments, comprise processor-implemented modules.
Similarly, the methods described herein may be at least partially processor-implemented. For example, at least some of the operations of a method may be performed by one or more processors or processor-implemented modules. The performance of some of the operations may be distributed among the one or more processors, not only residing within a single machine, but deployed across a number of machines. In some example embodiments, the processor or processors may be located in a single location (e.g., within a home environment, an office environment or as a server farm), while in other embodiments the processors may be distributed across a number of locations.
The one or more processors may also operate to support performance of the relevant operations in a “cloud computing” environment or as a “software as a service” (SaaS). For example, at least some of the operations may be performed by a group of computers (as examples of machines including processors), these operations being accessible via a network (e.g., the Internet) and via one or more appropriate interfaces (e.g., Application Program Interfaces (APIs).)
Example embodiments may be implemented in digital electronic circuitry, or in computer hardware, firmware, software, or in combinations thereof. Example embodiments may be implemented using a computer program product, e.g., a computer program tangibly embodied in an information carrier, e.g., in a machine-readable medium for execution by, or to control the operation of, data processing apparatus, e.g., a programmable processor, a computer, or multiple computers.
A computer program can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand-alone program or as a module, subroutine, or other unit suitable for use in a computing environment. A computer program can be deployed to be executed on one computer or on multiple computers at one site or distributed across multiple sites and interconnected by a communication network.
In example embodiments, operations may be performed by one or more programmable processors executing a computer program to perform functions by operating on input data and generating output. Method operations can also be performed by, and apparatus of example embodiments may be implemented as, special purpose logic circuitry, e.g., a field programmable gate array (FPGA) or an application-specific integrated circuit (ASIC).
The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In embodiments deploying a programmable computing system, it will be appreciated that that both hardware and software architectures require consideration. Specifically, it will be appreciated that the choice of whether to implement certain functionality in permanently configured hardware (e.g., an ASIC), in temporarily configured hardware a combination of software and a programmable processor), or a combination of permanently and temporarily configured hardware may be a design choice. Below are set out hardware (e.g., machine) and software architectures that may be deployed, in various example embodiments.
In some embodiments, the described methods may be implemented using one of a distributed or non-distributed software application designed under a three-tier architecture paradigm. Under this paradigm, various parts of computer code (or software) that instantiate or configure components or modules may be categorized as belonging to one or more of these three tiers. Some embodiments may include a first tier as an interface (e.g., an interface tier). Further, a second tier may be a logic (or application) tier that performs application processing of data inputted through the interface level. The logic tier may communicate the results of such processing to the interface tier, and/or to a backend, or storage tier. The processing performed by the logic tier may relate to certain rules, or processes that govern the software as a whole. A third, storage tier, may be a persistent storage medium, or a non-persistent storage medium. In some cases, one or more of these tiers may be collapsed into another, resulting in a two-tier architecture, or even a one-tier architecture. For example, the interface and logic tiers may be consolidated, or the logic and storage tiers may be consolidated, as in the case of a software application with an embedded database. The three-tier architecture may be implemented using one technology, or, a variety of technologies. The example three-tier architecture, and the technologies through which it is implemented, may be realized on one or more computer systems operating, for example, as a standalone system, or organized in a server-client, peer-to-peer, distributed or some other suitable configuration. Further, these three tiers may be distributed between more than one computer systems as various components.
Example embodiments my include the above described tiers, and processes or operations about constituting these tiers may be implemented as components. Common to many of these components is the ability to generate, use, and manipulate data. The components, and the functionality associated with each, may form part of standalone, client, server, or peer computer systems. The various components may be implemented by a computer system on an as-needed basis. These components may include software written in an object-oriented computer language such that a component oriented, or object-oriented programming technique can be implemented using a Visual Component Library (VCL), Component Library for Cross Platform (CLX), Java Beans (JB), Java Enterprise Beans (EJB), Component Object Model (COM), Distributed Component Object Model (DCOM), or other suitable technique.
Software for these components may further enable communicative coupling to other components (e.g., via various Application Programming interfaces (APIs)), and may be compiled into one complete server, client, and/or peer software application. Further, these APIs may be able to communicate through various distributed programming protocols as distributed computing components.
Some example embodiments may include remote procedure calls being used to implement one or more of the above described components across distributed programming environment as distributed computing components. For example, an interface component (e.g., an interface tier) may form part of a first computer system that is remotely located from a second computer system containing a logic component (e.g., a logic tier). These first and second computer systems may be configured in a standalone, server-client, peer-to-peer, or some other suitable configuration. Software for the components may be written using the above described object-oriented programming techniques, and can be written in the same programming language, or a different programming language. Various protocols may be implemented to enable these various components to communicate regardless of the programming language used to write these components. For example, a component written in C++ may be able to communicate with another component written in the Java programming language through utilizing a distributed computing protocol such as a Common Object Request Broker Architecture (CORBA), a Simple Object Access Protocol (SOAP), or some other suitable protocol. Some embodiments may include the use of one or more of these protocols with the various protocols outlined in the Open Systems Interconnection (OSI) model, or Transmission Control Protocol/Internet Protocol (TCP/IP) protocol stack model for defining the protocols used by a network to transmit data.
Example embodiments may use the OSI model or TCP/IP protocol stack model for defining the protocols used by a network to transmit data. In applying these models, a system of data transmission between a server and client, or between peer computer systems may for example include five layers comprising: an application layer, a transport layer, a network layer, a data link layer, and a physical layer. In the case of software, for instantiating or configuring components, having a three-tier architecture, the various tiers (e.g., the interface, logic, and storage tiers) reside on the application layer of the TCP/IP protocol stack. In an example implementation using the TCP/IP protocol stack model, data from an application residing at the application layer is loaded into the data load field of a TCP segment residing at the transport layer. This TCP segment also contains port information for a recipient software application residing remotely. This TCP segment is loaded into the data load field of an IP datagram residing at the network layer, Next, this IP datagram is loaded into a frame residing at the data link layer. This frame is then encoded at the physical layer, and the data transmitted over a network such as an internet, Local Area Network (LAN), Wide Area Network (WAN), or some other suitable network. In some cases, internet refers to a network of networks. These networks may use a variety of protocols for the exchange of data, including the aforementioned TCP/IP, and additionally ATM, SNA, SDI, or some other suitable protocol. These networks may be organized within a variety of topologies (e.g., a star topology), or structures.
Although an embodiment has been described with reference to specific example embodiments, it will be evident that various modifications and changes may be made to these embodiments without departing from the broader spirit and scope of the embodiment. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense. The accompanying drawings that form a part hereof, show by way of illustration, and not of limitation, specific embodiments in which the subject matter may be practiced. The embodiments illustrated are described in sufficient detail to enable those skilled in the art to practice the teachings disclosed herein. Other embodiments may be utilized and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. This Detailed Description, therefore, is not to be taken in a limiting sense, and the scope of various embodiments is defined only by the appended claims, along with the full range of equivalents to which such claims are entitled.
The tables 700 also include an items table 704 in which are maintained item records for goods and services that are available to be, or have been, transacted via the networked system 12. Each item record within the items table 704 may furthermore be linked to one or more user records within the user table 702, so as to associate a seller and one or more actual or potential buyers with each item record.
The items table 704 may be connected to an image table which contains images associated with the respective items or item listings in the items table 704, The image table 720 is in turn connected to an index data table 730 which contains index data as described in detail above.
A transaction table 706 contains a record for each transaction (e.g., a purchase or sale transaction) pertaining to items for which records exist within the items table 704.
An order table 708 is populated with order records, each order record being associated with an order. Each order, in turn, may be with respect to one or more transactions for which records exist within the transaction table 706.
Bid records within a bids table 710 each relate to a bid received at the networked system 12 in connection with an auction-format listing supported by an auction application 32. A feedback table 712 is utilized by one or more reputation applications 50, in one example embodiment, to construct and maintain reputation information concerning users. A history table 714 maintains a history of transactions to which a user has been a party. One or more attributes tables 716 record attribute information pertaining to items for which records exist within the items table 704. Considering only a single example of such an attribute, the attributes tables 716 may indicate a currency attribute associated with a particular item, the currency attribute identifying the currency of a price for the relevant item as specified in by a seller.
The example computer system 500 includes a processor 502 (e.g., a central processing unit (CPU) a graphics processing unit (GPU) or both), a main memory 504 and a static memory 506, which communicate with each other via a bus 508. The computer system 500 may further include a video display unit 510 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)). The computer system 500 also includes an alphanumeric input device 512 (e.g., a keyboard), a cursor control device 514 (e.g., a mouse), a disk drive unit 516, a signal generation device 518 (e.g., a speaker) and a network interface device 520.
The disk drive unit 516 includes a machine-readable medium 522 on which is stored one or more sets of instructions (e.g., software 524) embodying any one or more of the methodologies or functions described herein. The software 524 may also reside, completely, or at least partially, within the main memory 504 and/or within the processor 502 during execution thereof by the computer system 500, the main memory 504 and the processor 502 also constituting machine-readable media.
The software 524 may further be transmitted or received over a network 526 via the network interface device 520.
While the machine-readable medium 522 is shown in an example embodiment to be a single medium, the term “machine-readable medium” should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, and/or associated caches and servers) that store the one or more sets of instructions. The term “machine-readable medium” shall also be taken to include any medium that is capable of storing, encoding or carrying a set of instructions for execution by the machine and that cause the machine to perform any one or more of the methodologies described herein. The term “machine-readable medium” shall accordingly be taken to include, but not be limited to, solid-state memories, optical and magnetic media, and carrier wave signals.
Thus, a method and system to index images and to perform an image-based search in a network-based marketplace have been described. Although the present method and system have been described with reference to specific example embodiments, it will be evident that various modifications and changes may be made to these embodiments without departing from the broader spirit and scope of the application. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense.
The Abstract of the Disclosure is provided to comply with 37 CFR. §1.72(b), requiring an abstract that will allow the reader to quickly ascertain the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. In addition, in the foregoing Detailed Description, it can be seen that various features are grouped together in a single embodiment for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the claimed embodiments require more features than are expressly recited in each claim, Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed embodiment. Thus the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as a separate embodiment.
This application claims the benefit of priority of U.S. Provisional Application Ser. No. 61/318,771, filed on Mar. 29, 2010, which is incorporated herein by reference in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| 61318771 | Mar 2010 | US |
