Patent Application
20140063031
Some aspects relate to a method and system for providing image data to an image processing module.
Image or video processing systems can place heavy demands on memory systems. An image processing device is typically connected to a memory using a bus with a finite bandwidth. In many cases there will also be memory local to the image processing device, operating as a buffer or cache, which aims to reduce the amount of data being read from or written to the memory via the bus. Reading data from memory via a bus is relatively slow, and consumes more power, compared to reading the same data from local memory.
Some memory bandwidth is inevitable. A typical processing system must read each input image frame from memory via a bus, and write each output image frame to an memory via a bus. For High Definition (HD) video, a frame may be 1920 pixels wide and 1080 pixels high. With a typical 420 chrominance subsampling mode, the size of the frame is approximately 3.1 megabytes (MB). For an input frame rate of 24 frames per second (fps) the input data bandwidth alone is approximately 74 MB/sec.
The output video size and frame rate may differ from the input video size and frame rate, depending on the processing being applied. One possible processing algorithm is a high definition motion compensated frame rate converter, which may read input frames at 24 fps, and write output frames at, for example, 120 fps. In this case the output data bandwidth is in excess of 300 MB/sec, in addition to the input data bandwidth of 74 MB/sec described previously. This much data transfer is largely unavoidable.
An example frame rate conversion algorithm operates in two main phases. First, motion estimation computes a vector field representing the motion between a pair of consecutive input frames. Secondly, a picture building or interpolation phase constructs a number of output frames using pixels taken from the input frames, and placed in the output frames at positions determined by the motion vectors.
There are various approaches to motion estimation. A common one is to divide a frame into small rectangular blocks, and for each block, to search for a matching area of pixel data in an adjacent input frame. The search process typically requires the evaluation of each of a number of motion vector candidates, and the quality of the match is determined using a metric such as the sum of absolute differences (SAD) between the pixels in the block and the pixels from the adjacent frame. The positional offset between matching areas of pixels determines the motion vector for that block.
Motion vector candidates may be any vector within a search range. The search range surrounds the block, and its size determines the range of motions that can be detected and tracked. There are often several candidates with similar values, meaning that pixels from the adjacent frame must be read more than once. Therefore, it is appropriate to store the adjacent frame pixel data corresponding to the search range in a local buffer. Once the buffer is filled many vector candidates can be tested without consuming any additional memory bandwidth, however, the memory bandwidth consumption of filling the buffer must be considered.
In the picture building phase of the algorithm, pixel data corresponding to the selected motion vector is projected into its position in the output frame. Several alternative motion estimation results may be projected, with the pixel values in the output frame formed from a composite of each of them. The pixel data used in picture building is also read from the same pixel data buffer that supplies the pixels for motion estimation.
It is reasonable to expect that for motion estimation of a frame, every block of pixel data in the adjacent frame must be examined at least once. Most motion is approximately translational, and the vector field is relatively uniform. While there are exceptions where significant parts of the adjacent frame are never visited by the motion estimator, this cannot ever be relied upon to be the case. It can reasonably be said that a motion estimator is optimal in terms of memory bandwidth if each block of pixel data from the adjacent frame is read once, and once only.
The design of
A memory cache is not normally a suitable alternative to the pixel data buffer in this type of application, due to the time taken to retrieve data from memory via a bus in the event of a cache miss. In contrast, the pixel data buffer designs guarantee that data is available immediately. This is significant in a computationally intensive real-time application such as video processing.
As video frame sizes increase, the amount of storage required for the pixel data buffer increases proportionally. A high-definition (HD) image is 1920 by 1080 pixels. While there is no limit on the size of the motion in a video sequence, there are often practical constraints on the size of a motion estimator's search range. The vertical size of the search range, V, is of particular importance. For a motion estimator to track the majority of movements in HD video, the buffer may need to be several hundred pixels tall. If the pixel data buffer is the width of the screen and if the vertical search range, V, is 200 pixels, then storage for approximately 1920 by 200 pixels is required. (
Ultra High Definition Television (UHDTV) standards define new video formats with higher resolution than HD. 4K UHDTV is a format with frames twice the size of HD in each dimension, i.e. 3840 by 2160 pixels. The need to read the frames remains unavoidable, and so the memory bandwidth increases at least in proportion to the area of the frames, i.e. four times. Consequently, it is desirable to design the pixel data buffering to be as close to optimal as possible, adding no unnecessary memory bandwidth. This suggests the use of a pixel data buffer the full width of the screen. Typical motion vectors also scale in proportion to the frame size, so the search range dimensions will double. The full width pixel data buffer will therefore now be approximately 3840 by 400 pixels, four times the size, and therefore roughly four times the silicon area of the HD solution.
8K UHDTV frames are twice the size of 4K UHDTV in each dimension. This increases the bandwidth requirement to sixteen times that of HD video, and requires sixteen times the amount of pixel data buffer storage.
The applicant has appreciated that it would be desirable to reduce the amount of memory space required by image data buffers, while also reducing memory bandwidth consumption.
There is provided a method for providing image data to an image processing module, comprising the steps of: reading image data from memory into a down-scaler which down scales the image data to a first resolution; storing the first resolution image data in a first buffer; predicting a region of image data which the image processing module will request; storing image data corresponding to at least part of the predicted region of image data from memory into a second buffer, wherein the image data in the second buffer is at a higher resolution than the first resolution; receiving a request for image data from the image processing module; determining that image data corresponding to at least a portion of the requested image data is not in the second buffer; up-scaling image data from the first buffer to provide a substitute for the at least a portion of the requested image data at a resolution that at least matches the higher resolution; and providing the up-scaled image data to the image processing module.
Some implementations have an advantage that the amount of image data stored at the higher resolution in the second buffer can be small, without the risk that if image data corresponding to the requested image data is not in the second buffer image data will need to be fetched from memory, which can result in, for many applications, an unacceptable delay. If image data corresponding to requested image data is not in the second buffer, then image data from the first buffer is up-scaled to provide a substitute for the requested image data at a resolution that at least matches the higher resolution. This method has the disadvantage that, in some instances, lower quality image data may be provided to the image processing module. High frequency information is lost during the down-scaling process which can not be retrieved by subsequently up-scaling the image data. However, this method has the advantage that the memory space required by image data buffers can be reduced.
Even though image data for a region of the image may be stored in both the second and first buffers, since the second buffer (storing the higher resolution image data) can be small (because of the fall back of having the lower resolution image data in the first buffer), the total memory space required for both of the buffers can be less than that required if all of the image data stored in the first buffer were stored at the higher resolution.
This also means therefore that image data for a larger region of the image can be stored in the same amount of buffer memory. This has the advantage that, in some applications, memory bandwidth required to re-read image data which is not in a buffer can be reduced without increasing the amount of memory space required by the buffer.
In an implementation, the first buffer stores image data corresponding to the full width of an image frame.
Preferably, the image data in the second buffer is read into the second buffer from memory. However, in an alternative embodiment the step of storing image data corresponding to at least part of the predicted region of image data from memory into the second buffer comprises: reading image data from memory into a down-scaler which down-scales the image data to a second resolution; and storing the second resolution image data in the second buffer.
An embodiment of the invention further includes determining that image data corresponding to a portion of the requested image data is in the second buffer; and providing image data from the second buffer to the image processing module at a resolution at least equal to the higher resolution.
Image data may be arranged in each of the first and second buffers in rectangular blocks. Preferably, the size of the blocks in the first buffer is related to the size of the blocks in the second buffer by a scale factor corresponding to the difference between the first resolution and the higher resolution.
The second buffer may be implemented using a content addressable memory.
The region of image data which the image processing module will request may be predicted using data supplied by the image processing module or a further image processing module. For example, the region of image data which the image processing module will request may be predicted using motion vectors supplied by the image processing module or the further image processing module.
In an implementation, the first resolution image data is stored in the first buffer by: writing the first resolution image data to memory; and reading the first resolution image data from memory into the first buffer. Writing the first resolution image data to memory uses memory bandwidth. However, the applicant has appreciated that for applications where otherwise the image data would need to be re-read and down-scaled, writing the first resolution image data to memory may result in an overall memory bandwidth saving. The amount of memory bandwidth required to write and read the first resolution image data to and from memory may be less than the memory bandwidth required to read the image data at a higher resolution, and this saving will be increased if the image data needs to be read from memory multiple times. Further, this method enables the processing power required to down-scale image data to be reduced, since image data only needs to be down-scaled once.
In some aspects, the down-scaler produces residual image data by determining a difference between: image data at the higher resolution; and image data which has been up-scaled from the first resolution to the higher resolution.
In this embodiment, the image data corresponding to the at least part of the predicted region of image data in the second buffer may comprise residual image data. The residual image data has the same resolution as the higher resolution. However, the advantage of using residual image data is that residual image data can often be more easily and fully compressed than the higher resolution image data. This means that advantageously the memory bandwidth required to read and write the residual image data to and from memory may be less than the memory bandwidth required to read the higher resolution image data. Further, the memory space required by the second buffer may be reduced. The residual image data may be compressed using lossy or non-lossy compression.
In this embodiment, providing image data from the second buffer to the image processing module may comprise: up-scaling image data from the first buffer to at least match the higher resolution; combining residual image data from the second buffer with the up-scaled image data; and providing the result to the image processing module.
In some aspects, the method comprises determining the amount of high frequency information in a part of the image. This may be done by analysing the residual image data.
Image data corresponding to part of the predicted region of image data may not be stored into the second buffer, if it is determined that the amount of high frequency information in that part of the image is below a threshold. This enables memory bandwidth to be saved, and in addition frees up space in the second buffer for occupation by other data. This other data may be data that is yet to be stored, or thinking of the second buffer like a cache data which has already been stored (which might be useful in the future) which we can avoid evicting.
Image data corresponding to part of the predicted region of image data may also not be stored into the second buffer, if it is indicated that that part of the predicted region of image data will not be used by the image processing module. This may be the case, for example, if confidence in the predicted image data is low.
In a further preferred embodiment of the present invention, how much image data is stored from memory into the second buffer is determined in view of a memory bandwidth consumption rate requirement.
Preferably, the threshold is adjusted in view of the amount of high frequency information in different parts of the image and the memory bandwidth consumption rate requirement.
In a further aspect, first resolution image data may be provided from the first buffer to the image processing module or a further image processing module in response to a request from the image processing module or the further image processing module for image data at the first resolution. For example, the image processing module or the further image processing module may comprise a motion estimator which uses the image data at the first resolution.
The image processing module may be for example a picture builder of a frame rate convertor which outputs an image at at least the higher resolution based on the results of the motion estimator. In one embodiment, the picture builder of the frame rate convertor further comprises an analyser, which is provided with image data at the first resolution from the first buffer.
In a further aspect, a location of a potential discontinuity between the up-scaled image data from the first buffer and image data provided from the second buffer is identified. A blending unit combines image data from the second buffer with up-scaled image data from the first buffer at the vicinity of the discontinuity in order to reduce the visibility of the discontinuity.
In a further aspect, at least some chrominance image data corresponding to the at least part of the predicted region of image data is not stored into the second buffer. The applicant has appreciated that high frequency information is, in general, not as easily perceived in chrominance image data as in luminance image data. Therefore, not reading chrominance image data for one or more image pixels into the second buffer enables memory bandwidth to be conserved, whilst reducing any decrease in the quality of image data provided to the image processing module. It further enables space in the second buffer to be freed up for occupation by other data.
In some implementations, there is also provided a system for providing image data to an image processing module, comprising: a down-scaler configured to down-scale image data from memory to a first resolution; a first buffer arranged to store the first resolution image data; a second buffer arranged to store image data corresponding to at least part of a predicted region of image data, wherein the image data in the second buffer is at a higher resolution than the first resolution; and a processor configured to: read image data from memory into the down-scaler and store the first resolution image data in the first buffer; predict a region of image data which the image processing module will request; store image data corresponding to at least part of the predicted region of image data from memory into the second buffer; receive a request for image data from the image processing module; determine that image data corresponding to at least a portion of the requested region of image data is not stored in the second buffer; up-scale image data from the first buffer using an up-scaler to provide a substitute for the at least a portion of the requested region of image data at a resolution that at least matches the higher resolution; and provide the up-scaled image data to the image processing module.
Thus, in some implementations, the second buffer may operate a little like a cache, storing a number of blocks of pixel data at high resolution. The contents of the second (high resolution) buffer may be maintained by a buffer controller which may use prediction and rate control algorithms to provide the pixel data required by an image processing module, while also regulating the consumption of memory bandwidth.
Applications (such as the motion estimator and picture builder of a frame rate converter) may access data from either the first or second buffers. A motion estimator, for example, may be designed to operate entirely from low (first) resolution data from the first buffer. The picture builder should preferably operate with high resolution data, in order that the output frames maintain, wherever possible, the resolution of the input frames. Access to the second (high resolution) buffer is through the buffer controller, which supplies high resolution data where it is available in the high resolution buffer, and which supplies up-scaled data from the first (low resolution) buffer when high resolution data is not available.
Further features and advantages of implementations of the disclosure will become apparent from the following detailed description.
Example embodiments will now be described by way of example only with reference to the accompanying drawings in which:
The pixel buffering technique described here decouples on-chip memory size from input frame size while limiting memory bandwidth consumption.
This architecture provides a useful and flexible compromise between two extremes. The system of
The examples below are described in the context of a motion compensated frame rate converter. However, the skilled person will appreciate that the technique can be applied to a wide variety of image and video processing applications.
The term “application” is used here to describe a processing block or module that uses buffered pixel data. Applications are typically implemented as part of the same hardware device as the pixel buffering system and should not be confused with, for example, software applications running on a general purpose computer. The motion compensated frame rate converter is one example of an application that uses buffered pixel data, and it will be seen that this application can also be implemented as several applications, each using pixel data in a different way.
The down-scale process may use any one of the well-known techniques for image scaling. A reasonably high quality scaler, such as a bicubic type, is recommended, as pixels from the low resolution buffer may appear in the output image.
For some applications, the resolution of the down-scaled data is entirely sufficient. Such an application is shown as low resolution application 315. In the example of a frame rate converter it may be possible to run motion estimators as low resolution applications, as discussed below.
A buffer control unit 320 provides high resolution image data to a high resolution application 335. The high resolution application operates as if it has access to a full sized pixel data buffer containing high resolution pixel data, such as the one shown in
Buffer control unit 320 ensures that, wherever possible, the data requested by the high resolution application 335 is present in high resolution buffer 330.
In one example, pixel data storage in the high resolution buffer 330 is organised in rectangular blocks. The size of these storage blocks may be chosen for the best performance of the buffering system, and may differ from the block size, if any, used for processing by high resolution application 335. Division of the high resolution storage into blocks allows the management of the buffer contents at a convenient granularity, and for the request of regions of up-scaled data from the up-scaler 325 and low resolution buffer 310.
Similarly, low resolution buffer 310 may store pixel data in rectangular blocks. It may be particularly convenient, for example, to select a low resolution storage block size that is related to the high resolution storage block size by the scale factor used in the down-scaler 305 and up-scaler 325. In this way, one low resolution block corresponds to one high resolution block.
Memory caches are well known. A conventional memory cache receives requests for data from a particular memory address, and supplies that data to the device that made the request. The first time a piece of data is requested, it is fetched from memory via a bus. The data is duplicated in the cache, so that if requested a second time, the data can be supplied from the cache, rather than being fetched again from memory. This is known as a ‘cache hit’. The amount of storage in a cache is relatively small, so it is not generally possible for the cache to hold a duplicate copy of the entire working set of data. This is particularly true when the working set is large, as is the case with image or video data. Typically, a new piece of data being loaded into the cache will evict another piece of data. Should the evicted piece of data be requested again, it will need to be re-fetched from memory. The situation where a read request results in data being fetched from memory is known as a ‘cache miss’. The cache's size, mapping rules and replacement policy determine the behaviour of the cache, and can be selected for general purpose performance, or for performance in a particular application.
Pre-fetching is a technique used with caches when it is possible to predict that a piece of data will be requested from the cache in the near future. The pre-fetch transfers the data into the cache, such that a cache hit will occur when it is first requested. This may be appropriate in a video application, if, for example, blocks of pixels are being processed in a raster scan order. Shortly before processing of one block is complete, a pre-fetch instruction can be issued to fetch the data for the next block into the cache, such that the data for the next block is available immediately that the processing of the next block starts. Pre-fetching does not reduce bandwidth consumption, but does help to hide the latency (delay) associated with a cache miss. When pre-fetching is used, the possibility should be considered that the pre-fetched data will evict other data from the cache while it is still in use, causing additional cache misses.
The high resolution buffer 330 and controller 320 differ from a conventional cache in that a request for data that is not present in the high resolution buffer does not result in that data being fetched from memory. Instead, the data is provided from the low resolution buffer 310, via up-scaler 325. In this way, the latency of a memory access is always avoided.
Without a means such as the ‘cache miss’ to populate the high resolution buffer 330, some other mechanism must be provided. The buffer controller 320 performs this function in parallel with the provision of data to application 335.
The structure of a buffer control unit 320 is shown in
When the requested data is contained in the high resolution buffer, data is supplied via path 520 to multiplexer 525. If the requested data is not contained in the high resolution buffer a request is sent to the up-scaler (and from the up-scaler to the low resolution buffer) to supply up-scaled low resolution data via path 530. Alternatively, the request for data may be sent to high resolution buffer and up-scaler simultaneously, such that a greater part of the latency involved in upscaling low resolution data is hidden behind the time taken for the query of the high resolution buffer.
Optionally, the request for data may be sent to memory via a bus, and high resolution supplied to the multiplexer via path 535. Generally, the system supplies data through either low resolution or high resolution buffer paths to avoid the considerable latency and bandwidth consumption associated with an access to memory. Nevertheless, the option to supply data from memory may be useful in some applications.
Multiplexer 525 switches between the various data paths, under the control of read controller 500, so as to supply to appropriate data to the application, at output 540.
Pre-fetch controller 545 is responsible for reading data from memory and for storing it in the high resolution buffer. Ideally, all data will have been fetched and be available in the high resolution buffer before it is requested by an application. In practice, the availability of data will depend on the predictability of demand, the storage policy in the high resolution buffer (i.e. whether storing one piece of data causes the eviction of another), and other factors such as rate control applied to the memory interface. In general the pre-fetch control algorithm will depend upon the characteristics of the application.
The picture building phase of the frame rate converter is likely to run as a high resolution application. Motion vectors determined during a motion estimation phase are used to identify the locations in an input frame from which pixel data must be retrieved. The retrieved pixel data is then composited, possibly with other pixel data retrieved from other input frames, into an output frame for display.
Pre-fetch control for the picture building application is straightforward, if a motion estimation application runs sufficiently far in advance that motion vectors allow the exact data requirements of the picture builder to be known. In this case the pre-fetch control may comprise a FIFO (first in, first out) queue of the anticipated data requests, with data being supplied to the high resolution buffer just before it is requested by the picture builder.
If actual motion vectors are not available sufficiently far in advance, the pre-fetch controller may attempt to determine the requests that the picture builder is likely to make.
One example method of prediction is to assume that motion will be the same as it was at the same location during a previous frame interval. Moving objects usually continue to move, so this normally results in good prediction. An exception is at the edges of moving objects, where the vector corresponds to the motion of an object which has moved away. Motion compensation of the vector field may improve the prediction in these cases.
In many cases the dominant motion in a frame is due to motion of the camera, for example in the case of a horizontal pan. The vector representing this motion is known as a global motion vector, and may be derived from an average value of a vector field. Even where the scene contains moving objects, an average computed using a median produces a global motion vector which is often a good prediction for the motion vector in a majority of the blocks in a frame. A global motion vector therefore provides an alternative prediction which is suitable for use by the pre-fetch controller.
When using vectors from a previous frame interval, either individually, or as a global motion, the accuracy of the prediction relies upon motion that does not change significantly from one frame interval to the next. Fortunately, this is generally the case. As motion estimation proceeds in the current frame interval, those vectors may also be incorporated in the prediction, allowing small changes in motion to be accommodated.
When prediction fails the requested pixel data will not be available in the high resolution buffer, and so low resolution pixel data will be supplied instead. Prediction is most likely to fail where there are moving objects, or sudden changes in motion. Motion blur may soften the appearance of moving objects, and viewers are less able to track objects that move erratically. The supply of low resolution pixel data is therefore likely to coincide with areas of the image in which the high resolution content is either absent or not easily perceived. As a result, the visual impact of prediction failure is reduced.
In an ideal case, where prediction works well, the memory bandwidth consumption due to the supply of high resolution pixel data might be expected to approach the optimum case, where every pixel (or block of pixels) is read once. In practice, the blocks of pixels requested by the high resolution application are unlikely to be aligned with the storage block grid used in the high resolution buffer.
a shows common cases of a misaligned data requests. When block 700 is requested, pixels from four storage blocks 711, 712, 713, and 714 must be output. If the high resolution application processes blocks in a raster scan order, the caching effect of the high resolution buffer can reduce the memory bandwidth consumption from 4× to 2×. For example, when block 720 is requested, storage blocks 713 and 714 remain in the buffer from the previous access, and two additional blocks, 731 and 732 are required.
If resources permit, a larger high resolution buffer may allow blocks loaded during the processing of one row to remain available for reuse as the next row is being processed. For example, when block 740 is requested, storage blocks 712 and 714 remain in the buffer, and additional blocks 751 and 752 are required. When block 760 is then requested, blocks 714, 732, and 752 remain in the buffer, and one additional block, 753 is required. In this case, memory bandwidth consumption may approach 1×.
Alternatively, improvements in caching performance may be obtained without great increase in buffer size, by changing the block processing order from raster scan to a tiled layout, or a Morton order layout, if the application permits.
In a motion compensated system, block requests will be arranged in a packed grid (such as blocks 700, 720, 740, and 760) when the vector field is uniform. If the vector field is not uniform this may give rise to a sequence of block requests such as 740, 760, 780. While this sequence may be entirely predictable (given prior knowledge of the vector field), it is also likely to upset the caching performance of the buffer. For example, while storage blocks 732 and 792 may remain in the buffer, it is likely that blocks 731 and 791 have been discarded.
A motion estimator is a process which is well suited to operate as a low resolution application. Signal conditioning is a well-known technique to improve the performance of a motion estimator, and often includes low-pass filtering of the image data. The reasons that this is beneficial include noise reduction, and over-sensitivity of block matching measures such as the sum of absolute differences (SAD) to very high frequency detail. Consequently there is normally very little disadvantage, and probably some advantage, to running a motion estimator on data that has been down-scaled to a lower resolution.
Despite this, a motion estimator makes accesses to data that are reasonably predictable, and the degree of locality is such that the high resolution buffer can perform well if the choice is made to implement a motion estimator as a high resolution process.
In the motion estimator, the comparison of blocks of pixel data is typically the comparison of one grid-aligned block of pixels in one frame, with several non-candidate aligned candidate blocks in an adjacent frame. A number of non-aligned blocks are illustrated in
To the extent that candidate block positions can be predicted sufficiently far in advance, they provide useful input to the pre-fetch controller 545, since each candidate block position represents an area of pixel data that will be requested as the candidates are tested. The pre-fetch controller takes input from the application, via data path 550, which may include data of this type.
In many applications each frame of a video is processed several times. In the example of a motion estimator, a search takes place between a pair of frames. The first motion estimation takes place between frame 0 and frame 1. The second motion estimation takes place between frame 1 and frame 2, the third motion estimation between frame 2 and frame 3, and so on. With the exception of the first frame in the sequence, each frame is required by two motion estimations. More sophisticated motion estimators may perform searches over different intervals, and in different directions (e.g. searching for the contents of frame n in frame n−1 is not the same as searching for the contents of frame n−1 in frame n). As a result, it is possible that each frame of video will be read two, three, or more times. Since it is not normally feasible to store entire frames in buffer storage, memory bandwidth increases accordingly.
In an alternative embodiment, memory 400 contains storage for high resolution video frames 405, which are processed by a down-scaler 410. Low resolution buffer control 415 stores the down-scaled pixel data in the low resolution buffer, as before, and also returns the data to memory 420. When the same frame is processed for a second time, the low resolution buffer controller retrieves the low resolution frames 420 from memory, avoiding the need to down-scale the high resolution frame again.
If the size in memory of a high resolution frame is B, and a typical down-scale factor of 0.5 is used, then the size of the low resolution frame in memory is B/4. The first time a frame is processed, the memory bandwidth consumption is a read of B, plus a write of B/4. The second time the frame is processed the memory bandwidth consumption is a read of B/4. The total memory bandwidth consumption, for a frame processed n times, is (1+n/4) B, compared to nB for a system that does not store the down-scaled frames. There is therefore a bandwidth saving for any frame that is accessed more than once. Since the down-scaler 410 has to process fewer frames, there is also an opportunity to save silicon area or to reduce power consumption.
Down-scaler 410 may also produce a residual output. The residual is the difference between the high resolution pixel data, and the pixel data that will eventually be recreated when the low resolution pixel data is up-scaled. Residual data has the same resolution as the high resolution pixel data, but has the advantage of being relatively easy to compress. Residual encoder 425 compresses the residual data and stores it in memory at 430.
Buffer control unit 435 now maintains a buffer of encoded residual data 440, rather than the buffer of high resolution pixel data 330. When the high resolution application requests pixel data and the appropriate piece of residual data is available in residual buffer 440, the residual is decoded at 445, and added to the up-scaled low resolution data. This is equivalent to providing the high resolution data directly.
The compressibility of the residual data varies, between very little, in areas of considerable high frequency detail, to considerable, in areas where the high resolution frame has very little high frequency detail. In the worst case, the size of the residual data that must be fetched from memory is the same as if high resolution pixel data was being supplied. On average, the residual data is smaller, reducing the memory bandwidth further. Furthermore, compressed residual data occupies less space in residual buffer 440, increasing the effective size of the buffer, and meaning that there are fewer occasions when the buffer controller 435 will have to supply low resolution pixel data to the application.
Residual data may be stored in blocks, preferably the same size as the storage block size used by the high resolution buffer. Various methods may be appropriate for compression of the residual data, and such methods are well known from the fields of image and video encoding. Methods based on the discrete cosine transform (DCT), or on wavelets, are appropriate. With such methods it also becomes possible to consider lossy encoding of the residual data. Lossy encoding is used, for example, in JPEG image encoding, or in MPEG video encoding.
In the system described here, lossy encoding of residuals may further reduce memory bandwidth, and may permit an even greater number of residuals to be stored in the high resolution buffer than is possible with lossless residual encoding. The possible disadvantage is that the reconstruction of high resolution pixel data, from the up-scaled low resolution image and the residual, is no longer perfect. In many cases the degree of loss may be visually imperceptible, and in most cases, the reconstructed high resolution data will be preferred over the use of the up-scaled low resolution data, which itself represents a considerable degree of loss of information. Whether to use lossy residual compression, and the degree of compression used, will depend on the application.
Computing a residual also provides an opportunity to analyse the contents of the high resolution image. There are considerable benefits in performing this analysis, even in the system of
Where the residual for a block is sufficiently small that it can be determined that there will be no significant visual difference between high resolution and low resolution pixel data, the buffer controller may choose not to load high resolution data from memory, even when it is predicted that that data will be requested by a high resolution application. The absence of the data in the high resolution buffer will result in low resolution data being supplied instead, and a saving in bandwidth consumption will be achieved.
A similar benefit may be gained with information supplied by the applications.
In the frame rate converter example, the picture builder may analyse motion vector and pixel data as it constructs the composite image. Several situations may occur where the picture builder can supply application data that allows the pre-fetch controller to optimise bandwidth consumption.
As described previously, a typical picture builder creates its composite image using pixel data projected from several input frames, according to several different motion vector fields. The job of the picture builder is to resolve disagreements between the different projections in order to produce the best possible composite output. The output will typically be a weighted blend between some or all of the projections at any particular position in the image.
In one example the picture builder is split into two parts; an analysis stage, and a composition stage. The analysis stage now runs as a low resolution application, and the composition stage as a high resolution application.
Where the analysis stage determines that all, or a majority, of projections are consistent with each other, a blend is unnecessary, and any one of the consistent projections is sufficiently good to be used alone in the output image. In this case, the analysis stage can send application data to the pre-fetch controller indicating that only one of the projected images is required by the composition stage. The high resolution data corresponding to this projection will be fetched into the high resolution buffer, but data for other projections will not.
Where the degree of confidence in the correctness of the composite image is low various types of error concealment are used. One particular type of error concealment is to apply a low pass filter, or blur, to the output image. It is recognised that there is little point in reading high resolution data if the resolution is to be degraded by filtering. Therefore, the use of error concealment may be indicated by sending application data to the buffer controller. The buffer controller will avoid reading the high resolution data, resulting in a reduction in bandwidth consumption.
Similar savings may be made at any time that it can be determined that only a subset of projected images are required by the composition stage.
It will be clear that the above described examples are relevant to applications other than frame rate conversion, and that many algorithms will be able to supply application data that allows the pre-fetch controller to optimise memory bandwidth consumption in an appropriate way.
The examples above describe situations where the buffer controller avoids fetching certain blocks of high resolution data, based on information that indicates that the low resolution data will be sufficient. There may be other cases where high resolution data would be desirable, but for some reason is not available. Reasons may include insufficient space in the high resolution buffer, inaccurate prediction by the application or pre-fetch controller, or insufficient available memory bandwidth to supply the high resolution data.
In the latter case, the system may impose a limit on the amount of memory bandwidth consumed. In the description of
In one example, rate control may be applied to the high resolution data pre-fetch. For example, a system that would typically consume memory bandwidth of 2.2× the amount of requested data might be restricted to consume only 1.5×. The bandwidth restriction is applied by avoiding reading certain blocks of high resolution data, and instead supplying low resolution data to the applications instead.
Rate control algorithms are used in video coding where there is often a requirement to produce an encoded bit stream at a desired data rate. A video encoder controls the data rate by varying the degree of quantization used for each encoded block. In the system of
A typical rate control algorithm will achieve the target rate in the long term, while often allowing short term fluctuations above the target rate. In a frame based system, such as video processing, it is desirable to distribute the available bandwidth appropriately across the frame. The rate control should not waste bandwidth supplying high resolution data to the top part of a frame if this means that the lower part of the frame is starved of data. On the other hand, bandwidth should not be reserved excessively in anticipation of high resolution detail that does not exist. Analysis of the distribution of detail in the frame, such as has been described in conjunction with the computation of a residual, provides data that a rate controller can use to allocate bandwidth appropriately. Where application data also allows the pre-fetch controller to avoid reading high resolution data for some parts of the image, the application data may also be supplied to the rate controller and incorporated into the rate control calculations.
a shows a situation in which high resolution data is not fully available. Storage blocks shown shaded, such as 800, are blocks for which high resolution data is available, and blocks without shading, such as 805, are blocks for which high resolution data is not available. A request for pixel data block 810 can be satisfied entirely from the high resolution buffer. Pixel data block 815, however, requires data from storage block 805, which is not available in the high resolution buffer. The portion 820, at least, of the request must therefore be satisfied using low resolution data.
b shows a request for a further pixel data block 850, occurring some time after the request for block 810. High resolution data for storage block 800, previously available in the buffer, is no longer available, and so portion 855 of the request is satisfied using low resolution data.
When a block such as 815 or 825 is requested, the read controller has the option of providing only low resolution pixel data for the requested block, or, providing those high resolution pixels that are available and providing the remainder from the low resolution buffer. The decision may depend upon the proportion of available high resolution pixels, or their distribution within the block. In the case that a mixture of high and low resolution pixels are supplied the read control unit 500 of
The approach to resolution discontinuity is highly dependent on the application. It is always known, when data is requested, which pixels can be provided from the high resolution buffer, and which from the low resolution buffer. An application may then take into account data such as residual sizes, or other information obtained during image analysis, in order to determine the likely visibility of any discontinuity. One option is to conceal a discontinuity by the application of a smoothing post-process. This approach may be seen to be similar to de-blocking approaches already known in video decoders such as H.264. Alternatively, the supply of only low resolution data will avoid discontinuities within a block, but not necessarily at their edges. In
An alternative approach applies filtering or blending as the data is provided.
Video images are often represented using luminance and chrominance data. The luminance data contains a high resolution black and white image, and the chrominance data contains the additional data required to form a colour image. Chrominance data is commonly sub-sampled, such that it is at a lower resolution than the corresponding luminance data. It is known that the perception of detail and sharpness in an image comes primarily from the luminance data and that a considerable amount of information can be removed from the chrominance data before image degradation is easily perceived.
In another example, the buffer controller handles luminance and chrominance data separately. When a decision is made on whether to load data into the high resolution buffer, the prefetch controller may make a different decision for luminance data than it makes for the corresponding chrominance data. This decision may, for example, be based on the amount of high frequency detail in the luminance and chrominance data, on data from an application, or on the requirements of a rate controller. Requests for image data are handled according to whether luminance or chrominance data is available in the high resolution buffer, with any component that is unavailable being supplied from the low resolution buffer via the up-scaler. A system may be set up such that most chrominance data is supplied from the low resolution buffer, saving bandwidth and making more high resolution buffer space for luminance data. Alternatively, a system may supply all chrominance data from the low resolution buffer.
In the above examples, systems have been described in which image data exists at two resolutions. It is clear that the same techniques may be extended to any number of resolutions. For example, a system may add a medium resolution image buffer.
| Number | Date | Country | Kind |
|---|---|---|---|
| 1215870.5 | Sep 2012 | GB | national |
