Configuring spanning elements of a signature generator

Abstract

Systems, and method and computer readable media that store instructions for configuring spanning elements of a signature generator.

Description

BACKGROUND

Object detection has extensive usage in variety of applications, starting from security, sport events, automatic vehicles, and the like.

Vast amounts of media units are processed during object detection and their processing may require vast amounts of computational resources and memory resources.

There is a growing need to provide an efficient object detection method.

SUMMARY

There may be provided systems, methods and computer readable medium as illustrated in the specification.

BRIEF DESCRIPTION OF THE DRAWINGS

The embodiments of the disclosure will be understood and appreciated more fully from the following detailed description, taken in conjunction with the drawings in which:

FIG. 1 illustrates an example of a first dictionary;

FIG. 2 illustrates an example of a second dictionary;

FIG. 3 is an example of an image;

FIGS. 4 and 5 illustrate various original patches, sparse representations and reconstructed signal representing the original patches;

FIG. 6 is an example of a histogram;

FIG. 7 illustrates an example of augmented image;

FIG. 8 is an example of a graph that illustrates a frequency of appearance of various features in an image;

FIGS. 9 and 10 illustrate an example of an image and reconstructed images;

FIG. 11 illustrates an example of an image that includes groups of patches, an original image and a reconstructed image;

FIG. 12 illustrates an example of an image that includes groups of patches, an original image and a reconstructed image;

FIG. 13 illustrates an example of images;

FIG. 14 illustrates an example of a method;

FIG. 15 illustrates an example of a signature;

FIG. 16 illustrates an example of a dimension expansion process;

FIG. 17 illustrates an example of a clusters of a signatures matching process;

FIG. 18 illustrates a method;

FIG. 19 illustrates a method; and

FIG. 20 illustrates a system.

DESCRIPTION OF EXAMPLE EMBODIMENTS

In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be understood by those skilled in the art that the present invention may be practiced without these specific details. In other instances, well-known methods, procedures, and components have not been described in detail so as not to obscure the present invention.

The subject matter regarded as the invention is particularly pointed out and distinctly claimed in the concluding portion of the specification. The invention, however, both as to organization and method of operation, together with objects, features, and advantages thereof, may best be understood by reference to the following detailed description when read with the accompanying drawings.

It will be appreciated that for simplicity and clarity of illustration, elements shown in the figures have not necessarily been drawn to scale. For example, the dimensions of some of the elements may be exaggerated relative to other elements for clarity. Further, where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or analogous elements.

Because the illustrated embodiments of the present invention may for the most part, be implemented using electronic components and circuits known to those skilled in the art, details will not be explained in any greater extent than that considered necessary as illustrated above, for the understanding and appreciation of the underlying concepts of the present invention and in order not to obfuscate or distract from the teachings of the present invention.

Any reference in the specification to a method should be applied mutatis mutandis to a device or system capable of executing the method and/or to a non-transitory computer readable medium that stores instructions for executing the method.

Any reference in the specification to a system or device should be applied mutatis mutandis to a method that may be executed by the system, and/or may be applied mutatis mutandis to non-transitory computer readable medium that stores instructions executable by the system.

Any reference in the specification to a non-transitory computer readable medium should be applied mutatis mutandis to a device or system capable of executing instructions stored in the non-transitory computer readable medium and/or may be applied mutatis mutandis to a method for executing the instructions.

Any combination of any module or unit listed in any of the figures, any part of the specification and/or any claims may be provided.

The specification and/or drawings may refer to an image. An image is an example of a media unit. Any reference to an image may be applied mutatis mutandis to a media unit. A media unit may be an example of sensed information. Any reference to a media unit may be applied mutatis mutandis to any type of natural signal such as but not limited to signal generated by nature, signal representing human behavior, signal representing operations related to the stock market, a medical signal, financial series, geodetic signals, geophysical, chemical, molecular, textual and numerical signals, time series, and the like. Any reference to a media unit may be applied mutatis mutandis to sensed information. The sensed information may be of any kind and may be sensed by any type of sensors—such as a visual light camera, an audio sensor, a sensor that may sense infrared, radar imagery, ultrasound, electro-optics, radiography, LIDAR (light detection and ranging), etc. The sensing may include generating samples (for example, pixel, audio signals) that represent the signal that was transmitted, or otherwise reach the sensor.

The specification and/or drawings may refer to a spanning element. A spanning element may be implemented in software or hardware. Different spanning element of a certain iteration are configured to apply different mathematical functions on the input they receive. Non-limiting examples of the mathematical functions include filtering, although other functions may be applied.

The specification and/or drawings may refer to a concept structure. A concept structure may include one or more clusters. Each cluster may include signatures and related metadata. Each reference to one or more clusters may be applicable to a reference to a concept structure.

The specification and/or drawings may refer to a processor. The processor may be a processing circuitry. The processing circuitry may be implemented as a central processing unit (CPU), and/or one or more other integrated circuits such as application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), full-custom integrated circuits, etc., or a combination of such integrated circuits.

Any combination of any steps of any method illustrated in the specification and/or drawings may be provided.

Any combination of any subject matter of any of claims may be provided.

Any combinations of systems, units, components, processors, sensors, illustrated in the specification and/or drawings may be provided.

Any reference to an object may be applicable to a pattern. Accordingly—any reference to object detection is applicable mutatis mutandis to a pattern detection.

Sparse Decomposition of Images Via Unsupervised Dictionary Learning

You only look once (YOLO) is a highly popular object detection method and has multiple variants.

It has been found that YOLO and convolutional neural network based object detection provide object features that are heavily skewed towards the object types that they are trained on and do not give enough distinguishing power to patterns outside the labelled set.

Furthermore, the representation given by the higher layers of YOLO-type networks contain numerical values, both high and low, that are related to the object dimensions; picking only the highest features causes significant information loss.

Finally, convolutional CNNs filters do not give clear intuition of what the features at each layer look like and it is not straightforward to reconstruct the data given just activation patterns in a specific layer.

Sparse representation of image features based on dictionary learning an alternative that might help overcome these shortcomings.

Given a set of training signals {x₁, x₂, . . . x_n}, where x_i represents a patch of dimension m=h_p·w_p·c_p, the method attempts to find a dictionary D of dimension m by L, L being the number of dictionary entries, such that for each x_i, there exists a sparse representation α_i such that the number of non-zero entries in α_i, i.e., |α_i|₀<<L, while Dα_i reconstructs each x_i with high fidelity.

Algorithmically, this is equivalent to solving the following optimization problem:

$\min \frac{1}{n}\sum _{i=1}^n{D{\alpha }_i-x_i}_2^2$

Such that |α_i|₁<λ for all i

Wherein λ is a constraint that can be set by a user—for example —maximum five or ten per patch.

Efficient algorithms exist for performing stochastic gradient descent alternately on D and α, while convergence guarantees.

For example—there are shown examples of dictionaries learnt from various images.

Each one of FIGS. 1 and 2 illustrates dictionary learnt from sixty four training images [each training image is a Continental 1024×640 pixels, RGB] in one of the following settings:

FIG. 1—dictionary 9501—h_p=w_p=16, L=1024

FIG. 2—dictionary 9502—h_p=w_p=32, L=4096

Wherein h_p is the height of the patch and w_p is the width of the patch.

Sparse Feature Decomposition

Once a dictionary D is learned, it can be applied to new image patches to produce their respective sparse representation {α_i}. The same optimisation problem as before is solved while keeping D fixed. For the following example image 9503 (included in FIG. 3), the method applied the first dictionary with 1024 entries and patch dimension 16·16·3.

FIGS. 4 and 5 illustrate various original patches (9504(1), 9504(2) and 9504(3)) from image 9503, their sparse representation using D (9505(1), 9505(2) and 9505(3)) as well as the reconstructed signal (9506(1), 9506(2) and 9506(3)) representing the original patch.

The patches form a grid of patches and the location of each one of patches 9504(1), 9504(2) and 9504(3) is denoted by their row and column. First patch 9504(1) is denoted (15,15)—and is located at the 15'th row and the 15'th column in the grid of patches.

Second patch 9504(2) is denoted (15,45)—and is located at the 15'th row and the 45'th column in the grid of patches.

Third patch 9504(3) is denoted (25,5)—and is located at the 25'th row and the 5'th column in the grid of patches.

Fourth patch 9504(4) is denoted (35,45)—and is located at the 35'th row and the 45'th column in the grid of patches.

The sparse representation using D are illustrated by graphs or rather histograms that illustrate the amplitude of each dictionary element in the sparse representation. Most dictionary elements are zero.

FIG. 6 is an example of a histogram formed for image 9503. The y-axis if the number of patches and the x-axis is the number of active features (nonzero dictionary elements) per patch.

The histogram shows that in relation to image 9503 the majority of the patches can be represented by dozens of active features. In total, sparse decomposition yields 96608 non-zero entries, which translates to an average of about 38 active features per patch.

FIG. 7 illustrates an example of augmented image 9508 in which the number of active features (non-zero elements) per patch are overlaid on image 9503. This shows that various textures in the image requires different number of features to represent.

FIG. 8 is an example of a graph 9509 that illustrates the frequency of appearance of various features in image 9503.

Image Reconstruction

One advantage of the dictionary learning method is that back-forward reconstruction is as simple as a single instance of matrix multiplication, while CNN activations generally require some kind of non-linear optimization to traceback the input from the previous layer.

FIGS. 9 and 10 illustrates image 9503 and first, second, third, fourth and fifth reconstructed images 9511, 9512, 9513, 9514 and 9515 respectively.

The first till fifth reconstructed images differ from each other by the maximal number of features that may represent each patch.

In first reconstructed image 9511 the maximal active features per patch is 128. In second reconstructed image 9512 the maximal active features per patch is 64. In third reconstructed image 9513 the maximal active features per patch is 48. In fourth reconstructed image 9514 the maximal active features per patch is 32. In fifth reconstructed image 9515 the maximal active features per patch is 16.

The following shows the reconstructed image based on its sparse representation of no more than 100,000 floating-point numbers, with various degrees of clipping. In agreement to the distribution of active features (see previous histogram), high-fidelity reconstruction of the image is possible as long as each patch can be represented by at least a predefined number (for example—16, 32, 48, 64 and 128) active features. The exact threshold varies depending on the content of the image, but is expected to be lower in higher-dimensional representations.

The sparse representation illustrated above provides a viable alternative to convolutional neural network based object detection in terms of low-level feature extraction, with the advantage that it works on unlabeled, limited training data, and enables backward visualization.

An object detection method that uses the spare representation may include generating extract high-level, abstract features (for example high level features that are comparable to those given by YOLO/MSSD) by training additional layers of dictionaries on top of the already sparse features and taking into account more contextual information.

Similarity Pooling

The first layer of descriptors of an image was illustrated above and include assigning a sparse representation to each patch (for example of 16 by 16 pixels).

The patches may then be grouped to groups of patches—wherein each group may be calculated based on a similarity between patches.

Thus, after generating a first sparse layer (sp0) the patches are pooled according to their similarity instead of using a fixed 3×3 grid.

Since the sparse feature vectors in this case are sparse, the similarity (denoted S(a,b)) between two patches (patch a and patch b) may be calculated by a normalized dot product of their respective vectors (for example—even without mean subtraction):

$s\left(a,b\right)=\frac{v_a·v_b}{v_av_b}$

S(a,b) is compared to a threshold and once exceed patches a and b are deemed similar to each other.

In contrast to clustering in the higher layers, the similarity among patches that are located only within a certain radius (nearest-neighbors) are considered. Hence the computational complexity scales as O(n), where n is the number of patches, as opposed to O(n²) as in the case for all-to-all connectivity.

A fixed threshold is then applied to the resulting similarity.

Any connected subgraph is considered a single pooled patch for the next layer and is represented by an average feature vector.

FIG. 11 illustrates image 9521 that includes 1077 groups of patches (out of a total of 2560 patches per image), that were obtained using a pool radius of 1.5 patches and a similarity threshold of 0.5.

FIG. 11 illustrates that after the first layer of sparse decomposition, the features result in a low-level segmentation of the image. Homogeneous regions like road surface and sky are collected into single groups, whereas more complex objects remain fragmented.

FIG. 11 also illustrates original image 9522 and reconstructed image 9523. Reconstructed image 9523 is reconstructed using the resulting average feature vector that represents each region (sm0→sp0→img), with slight loss of quality.

Lower thresholds result in more aggressive pooling, and less information retention—as illustrated in images 9524, 9525 and 9526.

FIG. 12 illustrates image 9524 that includes 848 groups of patches (out of a total of 2560 patches per image), that were obtained using a pool radius of 1.5 patches and a similarity threshold of 0.3.

FIG. 12 also illustrates original image 9525 and reconstructed image 9526. Reconstructed image 9526 is reconstructed using the resulting average feature vector that represents each region (sm0→sp0→img), with slight loss of quality.

FIG. 13 illustrates image 9527 that includes 600 groups of patches (out of a total of 2560 patches per image), that were obtained using a pool radius of 1.5 patches and a similarity threshold of 0.3.

FIG. 13 also illustrates image 9528 that includes 1107 groups of patches (out of a total of 2560 patches per image), that were obtained using a pool radius of 1.5 patches and a similarity threshold of 0.3.

Configuring Spanning Elements

The analysis of content of a media unit may be executed by generating a signature of the media unit and by comparing the signature to reference signatures. The reference signatures may be arranged in one or more concept structures or may be arranged in any other manner. The signatures may be used for object detection or for any other use.

The signature may be generated by creating a multidimensional representation of the media unit. The multidimensional representation of the media unit may have a very large number of dimensions. The high number of dimensions may guarantee that the multidimensional representation of different media units that include different objects is sparse—and that object identifiers of different objects are distant from each other—thus improving the robustness of the signatures.

The generation of the signature is executed in an iterative manner that includes multiple iterations, each iteration may include an expansion operations that is followed by a merge operation. The expansion operation of an iteration is performed by spanning elements of that iteration.

FIG. 14 illustrates a method 5000 for generating a signature of a media unit.

Method 5000 may start by step 5010 of receiving or generating sensed information.

The sensed information may be a media unit of multiple objects.

Step 5010 may be followed by processing the media unit by performing multiple iterations, wherein at least some of the multiple iterations comprises applying, by spanning elements of the iteration, dimension expansion process that are followed by a merge operation.

The processing may include:

• • Step 5020 of performing a k'th iteration expansion process (k may be a variable that is used to track the number of iterations).
  • Step 5030 of performing a k'th iteration merge process.
  • Step 5040 of changing the value of k.
  • Step 5050 of checking if all required iterations were done—if so proceeding to step 5060 of completing the generation of the signature. Else—jumping to step 5020.

The output of step 5020 is a k'th iteration expansion results 5120.

The output of step 5030 is a k'th iteration merge results 5130.

For each iteration (except the first iteration)—the merge result of the previous iteration is an input to the current iteration expansion process.

The method may include a configuration step of configuring the spanning elements.

FIG. 15 is an example of a signature 6027 of a media unit that is an image 6000 and of an outcome 6013 of the last (K'th) iteration.

The image 6001 is virtually segments to segments 6000(i,k). The segments may be of the same shape and size but this is not necessarily so.

Outcome 6013 may be a tensor that includes a vector of values per each segment of the media unit. One or more objects may appear in a certain segment. For each object—an object identifier (of the signature) points to locations of significant values, within a certain vector associated with the certain segment.

For example—a top left segment (6001(1,1)) of the image may be represented in the outcome 6013 by a vector V(1,1) 6017(1,1) that has multiple values. The number of values per vector may exceed 100, 200, 500, 1000, and the like.

The significant values (for example—more than 10, 20, 30, 40 values, and/or more than 0.1%, 0.2%. 0.5%, 1%, 5% of all values of the vector and the like) may be selected. The significant values may have the values—but may eb selected in any other manner.

FIG. 15 illustrates a set of significant responses 6015(1,1) of vector V(1,1) 6017(1,1). The set includes five significant values (such as first significant value SV1(1,1) 6013(1,1,1), second significant value SV2(1,1), third significant value SV3(1,1), fourth significant value SV4(1,1), and fifth significant value SV5(1,1) 6013(1,1,5).

The image signature 6027 includes five indexes for the retrieval of the five significant values—first till fifth identifiers ID1-ID5 are indexes for retrieving the first till fifth significant values.

FIG. 16 illustrates various spanning elements 5061(1)-5061(3).

Each relevant spanning element may perform a spanning operation that includes assigning an output value that is indicative of an identity of the relevant spanning elements of the iteration. The output value may also be indicative of identities of previous relevant spanning elements (from previous iterations).

For example—assuming that spanning element number fifty is relevant and is associated with a unique set of values of eight and four—then the output value may reflect the numbers fifty, four and eight—for example one thousand multiplied by (fifty+forty) plus forty. Any other mapping function may be applied.

FIG. 16 also illustrates the steps executed by each spanning element:

• • Checking if the merge results are relevant to the spanning element (step 5091).
  • If-so—completing the spanning operation (step 5093).
  • If not—entering an idle state (step 5092).

FIG. 17 illustrates an example of a clusters of a signatures matching process.

It is assumed that there are multiple (M) cluster structures 4974(1)-4974(M). Each cluster structure includes cluster signatures, metadata regarding the cluster signatures.

For example—first cluster structure 4974(1) includes multiple (N1) signatures (referred to as cluster signatures CS) CS(1,1)-CS(1,N1) 4975(1,1)-4975(1,N1) and metadata 4976(1).

Yet for another example—M'th cluster structure 4974(M) includes multiple (N2) signatures (referred to as cluster signatures CS) CS(M,1)-CS(M,N2) 4975(M,1)-4975(M,N2) and metadata 4976(M).

FIG. 17 also illustrates a media unit signature 4972 that is compared to the signatures of the M cluster structures—from CS(1,1) 4975(1,1) till CS(M,N2) 4975(M,N2).

We assume that one or more cluster structures are matching cluster structures.

Once the matching cluster structures are found the method proceeds by generating shape information that is of higher accuracy then the compressed shape information.

For example—assuming that the matching signatures include CS(1,1) 2975(1,1), CS(2,5) 2975(2,5), CS(7,3) 2975(7,3) and CS(15,2) 2975(15,2).

The number of signatures per concept structure may change over time—for example due to cluster reduction attempts during which a CS is removed from the structure to provide a reduced cluster structure, the reduced structure structure is checked to determine that the reduced cluster signature may still identify objects that were associated with the (non-reduced) cluster signature—and if so the signature may be reduced from the cluster signature.

The signatures of each cluster structures are associated to each other, wherein the association may be based on similarity of signatures and/or based on association between metadata of the signatures.

Assuming that each cluster structure is associated with a unique object—then objects of a media unit may be identified by finding cluster structures that are associated with said objects. The finding of the matching cluster structures may include comparing a signature of the media unit to signatures of the cluster structures—and searching for one or more matching signature out of the cluster signatures.

Each cluster may be identified by a cluster identifier. The cluster identifier may differ by size (for example by number of object identifiers) than a signature. One or more cluster identifiers may identify the cluster. A cluster identifier may include identifiers that are shared between CSs, may include identifiers that appear (even if not shared) in a CS, and the like. The cluster identifier may be generated by applying any function on the CS of a cluster.

For example—assuming that a cluster identifier include identifiers that are shared between two or more CS of the cluster. If no identifier is shared between all CS of the cluster—then multiple cluster identifier may be required to represent a single cluster. The cluster identifier may include all (or at least some) of the signature that appear in one or more CSs.

FIG. 18 illustrates a method 9600 for configuring spanning elements of a signature generator.

Method 9600 may include the following steps:

• • Receiving test images 9602.
  • Generating, by applying a unsupervised dictionary learning process, sparse signatures of the test images; wherein a sparse signature of a test image comprises multiple vectors, each vector represents a test image segment; wherein each vector comprises more zero valued elements than non-zero valued elements; wherein each element represent appearances, in the test image segment, of a unique pattern associated with the element 9604.
  • Calculating occurrence information regarding an occurrence of the unique patterns in groups of test image segments 9606.
  • Selecting, based on the occurrence information, a set of combinations of unique patterns 9608.
  • Associating different members of the set to different spanning elements of the signature generator 9610.

The set of combinations may include most popular combinations of the unique patterns. The set may include a predefined number of most popular combinations, all combinations of the unique patterns that exceed a popularity threshold. The rule for determining the size and/or the popularity threshold may be fixed, or may change over time.

The method may include selecting the set of combinations based on popularity of combinations of the unique patterns and based on a distance between the combinations of the unique patterns. Any cost function may be selected for determining the set of combinations. For example—there may be applied a distance threshold or a distance range so that only unique patterns within the range (or above or below the threshold) may be elected—regardless of their popularity. Yet for another example—there may be applied a popularity threshold or a popularity range so that only unique patterns within the range (or above or below the threshold) may be elected—regardless of their distance. Yet for a further example—the selection may be based on a combination of popularity and distance.

The groups of test image segments may include test image segments that are adjacent to each other.

The groups of test image segments may be of a same size.

The calculating of the occurrence information may be related to groups of first size of test image segments; wherein the selecting may include selecting a first set of combinations of unique patterns; and wherein the associating may include associating different members of the first set to different first spanning elements of the signature generator.

The method may further include calculating second occurrence information regarding an occurrence of the unique patterns in second groups of test image segments; the second groups are of a second size that exceeds the first size; selecting, based on the second occurrence information, a second set of combinations of unique patterns; and associating different members of the second set to different second spanning elements of the signature generator. These steps may add a layer of representations that may be mapped to the first layer of representations and may for a cortex structure. The number of layers may exceed two.

Each spanning element may be configured to determine whether it is relevant to a signature generation process based on the unique combination of one or more object identifiers that are associated with the spanning element. Thus—when a spanning element receives as input (for example as a result of a previous merge iteration) the unique combination (either alone or on addition to other signatures)—spanning element is relevant and completes the expansion process. If a spanning element does not receive (at least) the unique combination—the spanning element does not output a expansion result.

FIG. 19 illustrates method 9420 of generating a signature of a sensed information unit.

Method 9420 may include the following steps:

• • Receiving or generating a sensed information unit 9422.
  • Calculating the signature of the sensed information unit by performing multiple iterations, wherein each iteration of at least some of the multiple iterations may include applying, by spanning elements related to the iteration, a dimension expansion process that may be followed by a merge operation. 9424

The spanning elements may be configured by executing method 9400.

Method 9420 may also include:

• • Finding at least one matching cluster, each matching cluster has a cluster signature that matches the signature of the sensed information 9426.
  • Determining that the sensed information unit includes at least one object that is associated with the at least one matching clusters 9428.

Method 9420 may be preceded by method 9600 of configuring the spanning elements utilized in step 9424.

FIG. 20 illustrates an example of a system capable of executing one or more of the mentioned above methods.

The system include various components, elements and/or units.

A component element and/or unit may be a processing circuitry may be implemented as a central processing unit (CPU), and/or one or more other integrated circuits such as application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), full-custom integrated circuits, etc., or a combination of such integrated circuits.

Alternatively, each component element and/or unit may implemented in hardware, firmware, or software that may be executed by a processing circuitry.

System 4900 may include sensing unit 4902, communication unit 4904, input 4911, processor 4950, and output 4919. The communication unit 4904 may include the input and/or the output.

Input and/or output may be any suitable communications component such as a network interface card, universal serial bus (USB) port, disk reader, modem or transceiver that may be operative to use protocols such as are known in the art to communicate either directly, or indirectly, with other elements of the system.

Processor 4950 may include at least some out of

• • Multiple spanning elements 4951(q).
  • Multiple merge elements 4952(r).
  • Signature generator 4958.

While the foregoing written description of the invention enables one of ordinary skill to make and use what is considered presently to be the best mode thereof, those of ordinary skill will understand and appreciate the existence of variations, combinations, and equivalents of the specific embodiment, method, and examples herein. The invention should therefore not be limited by the above described embodiment, method, and examples, but by all embodiments and methods within the scope and spirit of the invention as claimed.

In the foregoing specification, the invention has been described with reference to specific examples of embodiments of the invention. It will, however, be evident that various modifications and changes may be made therein without departing from the broader spirit and scope of the invention as set forth in the appended claims.

Moreover, the terms “front,” “back,” “top,” “bottom,” “over,” “under” and the like in the description and in the claims, if any, are used for descriptive purposes and not necessarily for describing permanent relative positions. It is understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the invention described herein are, for example, capable of operation in other orientations than those illustrated or otherwise described herein.

Furthermore, the terms “assert” or “set” and “negate” (or “deassert” or “clear”) are used herein when referring to the rendering of a signal, status bit, or similar apparatus into its logically true or logically false state, respectively. If the logically true state is a logic level one, the logically false state is a logic level zero. And if the logically true state is a logic level zero, the logically false state is a logic level one.

Those skilled in the art will recognize that the boundaries between logic blocks are merely illustrative and that alternative embodiments may merge logic blocks or circuit elements or impose an alternate decomposition of functionality upon various logic blocks or circuit elements. Thus, it is to be understood that the architectures depicted herein are merely exemplary, and that in fact many other architectures may be implemented which achieve the same functionality.

Any arrangement of components to achieve the same functionality is effectively “associated” such that the desired functionality is achieved. Hence, any two components herein combined to achieve a particular functionality may be seen as “associated with” each other such that the desired functionality is achieved, irrespective of architectures or intermedial components. Likewise, any two components so associated can also be viewed as being “operably connected,” or “operably coupled,” to each other to achieve the desired functionality.

Furthermore, those skilled in the art will recognize that boundaries between the above described operations merely illustrative. The multiple operations may be combined into a single operation, a single operation may be distributed in additional operations and operations may be executed at least partially overlapping in time. Moreover, alternative embodiments may include multiple instances of a particular operation, and the order of operations may be altered in various other embodiments.

Also for example, in one embodiment, the illustrated examples may be implemented as circuitry located on a single integrated circuit or within a same device. Alternatively, the examples may be implemented as any number of separate integrated circuits or separate devices interconnected with each other in a suitable manner.

However, other modifications, variations and alternatives are also possible. The specifications and drawings are, accordingly, to be regarded in an illustrative rather than in a restrictive sense.

In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word ‘comprising’ does not exclude the presence of other elements or steps then those listed in a claim. Furthermore, the terms “a” or “an,” as used herein, are defined as one or more than one. Also, the use of introductory phrases such as “at least one” and “one or more” in the claims should not be construed to imply that the introduction of another claim element by the indefinite articles “a” or “an” limits any particular claim containing such introduced claim element to inventions containing only one such element, even when the same claim includes the introductory phrases “one or more” or “at least one” and indefinite articles such as “a” or “an.” The same holds true for the use of definite articles. Unless stated otherwise, terms such as “first” and “second” are used to arbitrarily distinguish between the elements such terms describe. Thus, these terms are not necessarily intended to indicate temporal or other prioritization of such elements. The mere fact that certain measures are recited in mutually different claims does not indicate that a combination of these measures cannot be used to advantage.

While certain features of the invention have been illustrated and described herein, many modifications, substitutions, changes, and equivalents will now occur to those of ordinary skill in the art. It is, therefore, to be understood that the appended claims are intended to cover all such modifications and changes as fall within the true spirit of the invention.

It is appreciated that various features of the embodiments of the disclosure which are, for clarity, described in the contexts of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features of the embodiments of the disclosure which are, for brevity, described in the context of a single embodiment may also be provided separately or in any suitable sub-combination.

It will be appreciated by persons skilled in the art that the embodiments of the disclosure are not limited by what has been particularly shown and described hereinabove. Rather the scope of the embodiments of the disclosure is defined by the appended claims and equivalents thereof.

Claims

1. A method for configuring spanning elements of a signature generator, the method comprises: receiving test images;generating, by applying an unsupervised dictionary learning process, sparse signatures of the test images; wherein a sparse signature of a test image comprises multiple vectors, each vector represents a test image segment; wherein each vector comprises more zero valued elements than non-zero valued elements; wherein each element represent appearances, in the test image segment, of a unique pattern associated with the element;calculating occurrence information regarding an occurrence of the unique patterns in groups of test image segments;wherein the calculating of the occurrence information is related to groups of first size of test image segments;selecting, based on the occurrence information, a set of combinations of unique patterns; wherein the selecting comprises selecting a first set of combinations of unique patterns; andassociating different members of the set of combinations of the unique patterns to different spanning elements of the signature generator; wherein the associating comprises associating different members of the first set to different first spanning elements of the signature generator;calculating second occurrence information regarding an occurrence of the unique patterns in second groups of test image segments; the second groups are of a second size that exceeds the first size;selecting, based on the second occurrence information, a second set of combinations of unique patterns; andassociating different members of the second set to different second spanning elements of the signature generator.

2. The method according to claim 1 wherein the set of combinations comprises most popular combinations of the unique patterns.

3. The method according to claim 1 comprising selecting the set of combinations based on popularity of combinations of the unique patterns and based on a distance between the combinations of the unique patterns.

4. The method according to claim 1 wherein the test images comprise groups of test image segments, the groups of test image segments comprises test image segments that are adjacent to each other.

5. The method according to claim 1 wherein the test images comprise groups of test image segments, the groups of test image segments are of a same size.

6. The method according to claim 1 wherein the representations of the test image segments are signatures generated by the signature generator.

7. The method according to claim 1 wherein the representations of the test image segments differ from signatures generated by the signature generator.

8. The method according to claim 1 wherein the associating comprises gradually associating the different unique combinations to all spanning elements.

9. The method according to claim 1 comprising configuring the spanning elements based on a frequency of appearance of patterns in the test images.

10. The method according to claim 1 comprising assigning identifiers of a same object to a same spanning element.

11. The method according to claim 1 wherein each clusters of the representations is of at least a minimal predefined size and wherein a number of clusters of the representations is limited to a maximal predefined number.

12. The method according to claim 1 wherein the mapping comprises mapping each cluster identifier to set to a unique combination of one or more object identifiers, wherein at least one unique combination differs from a signature by number of object identifiers.

13. A method for generating a signature of a media unit, the method comprises: receiving or generating the media unit; andcalculating the signature of the media unit by performing multiple iterations, wherein each iteration of at least some of the multiple iterations comprises applying, by spanning elements related to the iteration, a dimension expansion process that is followed by a merge operation; wherein spanning elements related to the multiple iterations are configured by:receiving test media unit;generating representations of the test media units; wherein the representations are indicative of features of the test media units segments;finding a set of cluster identifiers that identify clusters of the representations of the test media units segments;mapping each decorrelated element of the set of cluster identifiers to a unique combination of one or more object identifier; andassociating different unique combinations to the spanning element of the signature generator.

14. A non-transitory computer readable medium for configuring spanning elements of a signature generator, the non-transitory computer readable medium stores instructions for: receiving test media units segments;generating representations of the test -media units segments; wherein the representations are indicative of features of the test media units segments;finding a set of cluster identifiers that identify clusters of the representations of the test media units segments;mapping each cluster identifier of the set of cluster identifiers to a unique combination of one or more object identifiers; andassociating different unique combinations of the one more object identifiers to the spanning element of the signature generator.

15. The non-transitory computer readable medium according to claim 14 wherein the finding of the set of the cluster identifiers is executed in an iterative manner, one subset of cluster identifiers after the other.

16. The non-transitory computer readable medium according to claim 14 wherein the finding of the set of cluster identifiers is executed in an iterative manner, one decorrelated element after the other.

17. The non-transitory computer readable medium according to claim 14 wherein the representations of the test media units segments are signatures generated by the signature generator.

18. The non-transitory computer readable medium according to claim 14 wherein the representations of the test media units segments differ from signatures generated by the signature generator.

19. The non-transitory computer readable medium according to claim 14 wherein the associating comprises gradually associating the different unique combinations to all spanning elements.

20. The non-transitory computer readable medium according to claim 14 that stores instructions for configuring the spanning elements based on a frequency of appearance of patterns in the test images.

21. The non-transitory computer readable medium according to claim 14 that stores instructions for assigning identifiers of a same object to a same spanning element.

22. The non-transitory computer readable medium according to claim 14 that stores instructions for searching for candidate cluster identifiers in a random manner and finding, out of candidate decorrelated elements, the set of decorrelated elements.

23. The non-transitory computer readable medium according to claim 14 wherein each clusters of the representations is of at least a minimal predefined size and wherein a number of clusters of the representations is limited to a maximal predefined number.

24. The non-transitory computer readable medium according to claim 14 wherein the mapping comprises mapping each cluster identifier to set to a unique combination of one or more object identifiers, wherein at least one unique combination differs from a signature by number of object identifiers.