Patent Application
20120183050
The present invention relates to the field of image processing. More specifically, the present invention relates to using a parametric loop filter in image processing.
In image processing, an adaptive loop filter (ALF) is performed to remove or reduce quantization noise and artifacts introduced during compression through wiener filtering as shown in
In quad-tree based on/off signaling, for each block (except those in the lowest layer), 1 bit is sent to specify if it is partitioned or not (partitioning signaling). For each non partitioned block (blocks in the lowest layer are always not partitioned), 1 additional bit is sent to specify if the ALF is applied to the block (on/off switch signaling). Given the Wiener filter, the quad-tree structure is optimized using bottom up recursive decision as shown in
A Wiener filter is trained online based on the statistics of the pixels that have the ALF on. Training is done by the encoder. In some embodiments, each frame has a Wiener filter, of which the filter coefficients are sent to the decoder. Training of Wiener filters needs to know the on/off information of the pixels. However, the on-off decision also requires knowledge of the Wiener filter.
The iteration decision includes training an initial Wiener filter (e.g. based on all pixels), deciding which position to place an on/off switch based on the current Wiener filter, updating the Wiener filter based on all “on” pixels, and performing deciding and updating multiple times (e.g. until conversion).
There are many drawbacks of ALF. In the encoder, there is high computation complexity since there are multiple passes of optimization in filter training and on/off switching. The encoder requires additional delay of one frame (without ALF, the delay is in the order of lines). The encoder stores/retrieves the cross- and auto-correlation matrices for each leaf node multiple times meaning significant memory accesses. The bitstream includes overhead bits signaling the filter coefficients. In the decoder, the Wiener filter coefficients are variable, and multiplication between two variables is expensive compared to multiplying a variable by a fixed number which is able to be replaced with several bit-shifts and additions. Additionally, the hardware implementation of these ALF techniques is very costly, particularly for real-time encoding, because training of Wiener filters has to be done after the entire frame is encoded, and the results of ALF are needed for motion estimation of the next frame. Therefore, the entire pipeline is prolonged, which requires a higher clock rate and more power consumption to meet the real-time constraint.
A parametric loop filter uses a set of fixed filters to remove or reduce noise and artifacts introduced during video coding. The parametric loop filter utilizes much less computation, delay and memory access at the encoder. Fixed coefficients allow fast implementation of filtering at the decoder. To specify a 2-D ideal low-pass filter, the following information is used: direction, bandwidth along the direction (bw//) and bandwidth perpendicular to the direction (bw⊥).
In one aspect, a parametric loop filter for filtering quantization noise introduced during compression comprises a set of fixed filters and a selecting mechanism for selecting one of the fixed filters to filter a block. Each filter of the set of fixed filters is indexed by one or more parameters. The one or more parameters identify the filter to an encoder and a decoder. The one or more parameters are predictive coded and sent to a decoder. The one or more parameters comprise a direction, a first bandwidth along the direction and a second bandwidth perpendicular to the direction. The direction is derived using an edge detection scheme or predicted from neighboring blocks. The first bandwidth and the second bandwidth are predicted using neighboring blocks or derived based on local characteristics of the block. A filter of the set of fixed filters is selected by an encoder and a uniquely identifiable parameter associated with the filter is transmitted to a decoder. The filter receives filter coefficients, wherein the filter coefficients are determined using known filters obtained by offline training. Each filter of the set of fixed filters is a near ideal or ideal low-pass filter. The parametric loop filter is implemented in hardware. The parametric loop filter is stored in a memory and processed by a processor. The filter is combined with online training to further improve performance, wherein an encoder and a decoder are aware of a default set of filters, and the encoder has the option to decide if a certain filter should be updated to the online trained filter: for every predefined filter for a certain direction and bandwidth, if the filter is updated, the encoder sends a 1 to the decoder, followed by the coefficients of the new filter, otherwise a 0 is sent, and wherein the decision of updating or not is made per frame or every few frames. The filter is implemented by a device selected from the group consisting of a personal computer, a laptop computer, a computer workstation, a server, a mainframe computer, a handheld computer, a personal digital assistant, a cellular/mobile telephone, a smart appliance, a gaming console, a digital camera, a digital camcorder, a camera phone, an iPod®/iPhone/iPad, a video player, a DVD writer/player, a Blu-ray® writer/player, a television and a home entertainment system.
In another aspect, a method of filtering a quantization noise introduced during compression using a parametric loop filter comprises determining which filter to use from a set of fixed filters and using a selected filter to filter data. Each filter of the set of fixed filters is indexed by one or more parameters. The one or more parameters identify the filter to an encoder and a decoder. The one or more parameters are predictive coded and sent to a decoder. The one or more parameters comprise a direction, a first bandwidth along the direction and a second bandwidth perpendicular to the direction. The direction is derived using an edge detection scheme or predicted from neighboring blocks. The first bandwidth and the second bandwidth are predicted using neighboring blocks or derived based on local characteristics of a block. The selected filter is selected by an encoder and a uniquely identifiable parameter associated with the filter is transmitted to a decoder. The filter receives filter coefficients, wherein the filter coefficients are determined using known filters obtained by offline training. Each filter of the set of fixed filters is a near ideal or ideal low-pass filter. The parametric loop filter is implemented in hardware. The parametric loop filter is stored in a memory and processed by a processor. The filter is combined with online training to further improve performance, wherein an encoder and a decoder are aware of a default set of filters, and the encoder has the option to decide if a certain filter should be updated to the online trained filter: for every predefined filter for a certain direction and bandwidth, if the filter is updated, the encoder sends a 1 to the decoder, followed by the coefficients of the new filter, otherwise a 0 is sent, and wherein the decision of updating or not is made per frame or every few frames. The filter is implemented by a device selected from the group consisting of a personal computer, a laptop computer, a computer workstation, a server, a mainframe computer, a handheld computer, a personal digital assistant, a cellular/mobile telephone, a smart appliance, a gaming console, a digital camera, a digital camcorder, a camera phone, an iPod®/iPhone/iPad, a video player, a DVD writer/player, a Blu-ray® writer/player, a television and a home entertainment system.
In yet another aspect, an apparatus for filtering a quantization noise introduced during compression comprises a direction module for determining a direction, a parallel bandwidth module to determine a first bandwidth along the direction, a perpendicular bandwidth module to determine a second bandwidth perpendicular to the direction, a selection module for selecting a filter from a set of fixed filters based on the direction, the first bandwidth and the second bandwidth and a filtering module for using the selected filter to filter noise in a block. Each filter of the set of fixed filters is indexed by the direction, the first bandwidth and the second bandwidth. The direction is derived using an edge detection scheme or predicted from neighboring blocks. The first bandwidth and the second bandwidth are predicted using neighboring blocks or derived based on local characteristics of a block. A uniquely identifiable parameter associated with the selected filter is transmitted to a decoder. The filter receives filter coefficients, wherein the filter coefficients are determined using known filters obtained by offline training. Each filter of the set of fixed filters is a near ideal or ideal low-pass filter. The apparatus is combined with online training to further improve performance, wherein an encoder and a decoder are aware of a default set of filters, and the encoder has the option to decide if a certain filter should be updated to the online trained filter: for every predefined filter for a certain direction and bandwidth, if the filter is updated, the encoder sends a 1 to the decoder, followed by the coefficients of the new filter, otherwise a 0 is sent, and wherein the decision of updating or not is made per frame or every few frames. The apparatus is selected from the group consisting of a personal computer, a laptop computer, a computer workstation, a server, a mainframe computer, a handheld computer, a personal digital assistant, a cellular/mobile telephone, a smart appliance, a gaming console, a digital camera, a digital camcorder, a camera phone, an iPod®/iPhone/iPad, a video player, a DVD writer/player, a Blu-ray® writer/player, a television and a home entertainment system.
In yet another aspect, an apparatus comprises a memory for storing an application, the application for determining which filter to use from a set of fixed filters indexed by a set of parameters including a direction, a first bandwidth along the direction and a second bandwidth perpendicular to the direction and using a selected filter to filter data and a processing component coupled to the memory, the processing component configured for processing the application. The set of parameters identify the filter to an encoder and a decoder. The set of parameters are predictive coded and sent to a decoder. The direction is derived using an edge detection scheme or predicted using neighboring blocks. The first bandwidth and the second bandwidth are predicted using neighboring blocks or derived based on local characteristics of a block. A uniquely identifiable parameter associated with the selected filter is transmitted to a decoder. Each filter of the set of fixed filters is a near ideal or ideal low-pass filter. The apparatus is combined with online training to further improve performance, wherein an encoder and a decoder are aware of a default set of filters, and the encoder has the option to decide if a certain filter should be updated to the online trained filter: for every predefined filter for a certain direction and bandwidth, if the filter is updated, the encoder sends a 1 to the decoder, followed by the coefficients of the new filter, otherwise a 0 is sent, and wherein the decision of updating or not is made per frame or every few frames.
When encoding a video, a picture is divided into several macroblocks (or coding units, coding tree units). After encoding the macroblocks, the encoder selects an encoding mode according to a desired bit rate and distortion and performs encoding in the selected encoding mode.
A parametric loop filter (also referred to as a parametric adaptive loop filter) uses a set of fixed filters which utilize much less computation, delay and memory access at the encoder. Fixed coefficients allow fast implementation of filtering at the decoder. Each filter is indexed by several parameters (e.g. given a set of parameters, p1, p2, . . . , pn, the filter is able to be uniquely identified by both encoder and decoder). The parameters that specify the filter of a block are predictive coded and sent to the decoder. The encoder determines the best filter in the set of filters for each block of an image. The selection of the best filter at the encoder is able to be performed in a single pass processing for each large coding unit or block. In some embodiments, only the parameter (also referred to as an index) of the filter is transmitted to the decoder. In some embodiments, the parameter is uniquely identifiable.
Using low-pass filters as candidate filters, if the compression is modeled as introducing additive noise to the original image, then ideal low-pass filters are a good approximation of Wiener filters. To specify a 2-D ideal low-pass filter, the following information is used: direction, bandwidth along the direction (bw//) and bandwidth perpendicular to the direction (bw⊥). Examples of typical filters are illustrated in
For blocks where the direction information is already available (e.g. intra-coded blocks that have a prediction direction), that direction is used as the prediction. For blocks where the direction information is not available, the direction is able to be derived using edge detection based schemes, since the decoder has access to the decoded current block. Alternatively, the direction information is able to be sent using Pulse Code Modulation (PCM), or based on a prediction from neighboring blocks. Bandwidths are able to be predicted using neighboring blocks or based on local characteristics of the block (e.g. the gradient strengths along or across the specified direction). The difference between the true parameter and its prediction is binarized using truncated unary code and coded using Context-Adaptive Binary Arithmetic Coding (CABAC).
For blocks where the direction information is already available (e.g. intra-coded blocks that have a prediction direction), that direction is used as the prediction. For blocks where the direction information is not available, the direction is able to be derived using edge detection based schemes, since the decoder has access to the decoded current block. Alternatively, the direction information is able to be sent using PCM, or based on a prediction from neighboring blocks. Bandwidths are able to be predicted using neighboring blocks or based on local characteristics of the block (e.g. the gradient strengths along or across the specified direction). The difference between the true parameter and its prediction is binarized using truncated unary code and coded CABAC.
In some embodiments, the parametric loop filter application(s) 630 include several applications and/or modules. Modules include a selection module to select a filter from a set of fixed filters to use for a block of data and a filtering module for using the selected filter to filter the block. In some embodiments additional modules include a direction module for determining a direction, a parallel bandwidth module to determine a bandwidth along the direction (bw//) and a perpendicular bandwidth module to determine a bandwidth perpendicular to the direction (bw⊥). In some embodiments, modules include one or more sub-modules as well. In some embodiments, fewer or additional modules are able to be included.
Examples of suitable computing devices include a personal computer, a laptop computer, a computer workstation, a server, a mainframe computer, a handheld computer, a personal digital assistant, a cellular/mobile telephone, a smart appliance, a gaming console, a digital camera, a digital camcorder, a camera phone, an iPod®/iPhone/iPad, a video player, a DVD writer/player, a Blu-ray® writer/player, a television, a home entertainment system or any other suitable computing device.
In some embodiments, parametric adaptive loop filters (PALF) are able to be used for high efficiency video coding. For luma, M×N fixed filters are pre-trained offline with the combination of M directions and N bandwidths plus one additional all pass filter (of which the impulse response is a delta function). Each largest coding unit (LCU) decides on a maxDepth used for signaling of the PALF parameters. This maxDepth is signaled at each LCU of the decoder. PALF parameters signaled at each CU (depth<=maxDepth) includes a direction (e.g. 1 . . . M) and filter strength (bandwidth) along the direction (e.g. 0 . . . N) where filter strength of 0 means no filtering (in which case the direction is not sent). The filter strength (bandwidth) perpendicular to the direction is assumed to be 0 (all-pass). The parameters of direction and strength are able to be predicted from neighboring CUs and coded using variable length codes. PALF parameters are decided once for each LCU, then the LCU is filtered according to the parameters and no further access is needed of that LCU. In some embodiments, where the filters are trained as 9×9 in size, this yields a maximum encoder delay of 1 LCU+4 lines.
In some embodiments, 8 directions and 4 bandwidths are used. The 8 directions are VERTICAL, HORIZONTAL, DIAG_DOWN_LEFT, DIAG_DOWN_RIGHT, VER_RIGHT, HOR_DOWN, VER_LEFT and HOR_UP (same as the directions in AVC intra prediction), and are indexed using 1˜8 respectively. In coding of the filter bandwidth, if the filter strength level n (0≦n≦4) of the current CU is equal to the predicted strength level which is obtained from the left and top CUs, then only a 1 is transmitted. Otherwise, a 0 is transmitted and the actual level n is signaled utilizing a fixed length code. Except for the case of the all-pass filter, signaling of the direction of the filter follows next. The prediction of the strength level is equal to the minimum strength level between left and top CUs. The method of coding direction is similar, but if the predicted direction is different from the true direction, a variable length code of 2 bits or 3 bits needs to be sent in addition to identify from the rest 7.
The encoder optimization includes finding the best max depth Dmax for each LCU and the best filter indices (m, n) for all CUs within the LCU. This is accomplished by finding the Sum of Squared Error (SSE) between the original block and the filtered version of the reconstructed block. The SSE associated with all possible filters is computed in the case of an exhaustive search method. However, many fast search algorithms are able to be used to avoid an exhaustive search. Starting from the first CU, the filter with the minimum Rate-Distortion (RD) cost is selected for each CU. To further reduce the computational cost of the encoder, the signaling cost of each filter is approximated when an arithmetic entropy coding engine is used.
For chroma, in some embodiments, chroma filters are 5×5 for YUV420 sequences. Signaling of the filter parameters for the chroma channels is similar to that of luma, except that the maximum depth Dmax is typically of a smaller value. Both chroma components use the same filter strength and direction which minimizes the sum of their RD costs. For the filter direction of the chroma components, a flag is sent to signal whether the direction is the same as luma or not. If the direction is different, then, it is signaled similarly to the luma directions. Otherwise, no extra bits will be necessary, and the directions of corresponding luma blocks are used for chroma components.
In some applications, the encoder is not sensitive to delay and memory access. In such applications, PALF is able to be combined with online training to further improve performance. Both the encoder and the decoder are aware of the default set of filters (e.g. offline trained). The encoder has the option to decide if a certain filter should be updated to the online trained filter. For every predefined filter (for a certain direction and bandwidth, if a filter is updated, the encoder sends a “1” to the decoder, followed by the coefficients of the new filter, otherwise a “0” is sent. The decision of updating or not is able to be made per frame or every few frames.
To utilize a parametric loop filter, a user acquires a video/image such as on a digital camcorder, and while or after the video is acquired, or when displaying the video, the parametric loop filter is automatically used for filtering the image, so that noise in the image is removed. The parametric loop filter is able to be implemented automatically without user involvement.
In operation, the parametric loop filter is able to filter noise from data in a much more efficient manner than a Wiener filter, particularly at low bit-rate applications, where there is not enough bit budget to signal Wiener filter coefficients. The parametric loop filter utilizes fewer computations, less delay and less memory access at the encoder. The delay is less because the encoder only processes each block once, and after the last block, there are no additional access to current frame. Since there is less memory access required, there are low memory and memory bus requirements. The fixed filters are hardware friendly and fast. Fewer computations also results in low power usage.
In contrast, the traditional ALF has high complexity with many operations needed to find filter coefficients. Variable filter coefficients are costly, and many pass processing at the encoder increases delay and requires high bus bandwidth which is not hardware friendly.
a. a set of fixed filters; and
b. a selecting mechanism for selecting one of the fixed filters to filter a block.
a. determining which filter to use from a set of fixed filters; and
b. using a selected filter to filter data.
The present invention has been described in terms of specific embodiments incorporating details to facilitate the understanding of principles of construction and operation of the invention. Such reference herein to specific embodiments and details thereof is not intended to limit the scope of the claims appended hereto. It will be readily apparent to one skilled in the art that other various modifications may be made in the embodiment chosen for illustration without departing from the spirit and scope of the invention as defined by the claims.
This application claims priority under 35 U.S.C. §119(e) of the U.S. Provisional Patent Application Ser. No. 61/433,026, filed Jan. 14, 2011 and titled, PARAMETRIC LOOP FILTER.” The Provisional Patent Application Ser. No. 61/433,026, filed Jan. 14, 2011 and titled, “PARAMETRIC LOOP FILTER” is also hereby incorporated by reference in its entirety for all purposes.
| Number | Date | Country | |
|---|---|---|---|
| 61433026 | Jan 2011 | US |
