Patent Application
20140118460
In the past, the technique known as “super resolution” has been used in satellite imaging to boost the resolution of the captured image beyond the intrinsic resolution of the image capture element. This can be achieved if the satellite (or some component of it) moves by an amount corresponding to a fraction of a pixel, so as to capture samples that overlap spatially. In the region of overlap, a higher resolution sample can be generated by extrapolating between the values of the two or more lower resolution samples that overlap that region, e.g. by taking an average. The higher resolution sample size is that of the overlapping region, and the value of the higher resolution sample is the extrapolated value.
The idea is illustrated schematically in
More recently the concept of super resolution has been proposed for use in video coding. There are two potential applications of this. The first is similar to the scenario described above—if the user's camera physically shifts between frames by an amount corresponding to a non-integer number of pixels (e.g. because it is a handheld camera), and this motion can be detected (e.g. using a motion estimation algorithm), then it is possible to create an image with a higher resolution than the intrinsic resolution of the camera's image capture element by extrapolating between pixel samples where the pixels of the two frames partially overlap.
The second potential application is to deliberately lower the resolution of each frame and introduce an artificial shift between frames (as opposed to a shift due to actual motion of the camera). This enables the bit rate per frame to be lowered. Referring to
Embodiments of the present invention receive as an input a video signal comprising a plurality of frames, each comprising a plurality of higher resolution samples. For each respective one of the frames, multiple different projections of the respective frame are generated. Each projection comprises a plurality of lower resolution samples representing the respective frame at a lower resolution, wherein the lower resolution samples of the different projections represent different but overlapping groups of the higher resolution samples of the respective frame. The video signal is encoded by encoding the projections of each of the respective frames.
Further embodiments of the present invention receive a video signal comprising a plurality of frames, each frame comprising multiple different projections wherein each projection comprises a plurality of lower resolution samples. The lower resolution samples of the different projections represent different but overlapping portions of the respective frame. The video signal is decoded by decoding the projections of each of the respective frames. Higher resolution samples are generated representing each of the respective frames at a higher resolution. This is done by, for each higher resolution sample thus generated, forming the higher resolution sample from a region of overlap between ones of the lower resolution samples from the different projections of the respective frame. The video signal is output to a screen at the higher resolution following generation from the projection.
Various embodiments may be embodied as an encoding system, decoding system, or computer program code to be run at the encoder or decoder side, or may be practiced as a method. The computer program may be embodied on a computer-readable medium. The computer-readable may be a tangible, computer-readable storage medium.
For a better understanding of the various embodiments and to show how they may be put into effect, reference is made by way of example to the accompanying drawings in which:
The original use of super resolution was to artificially boost the resolution of a captured imaged beyond the intrinsic resolution of the capturing apparatus. As discussed, the idea was later proposed for use in video transmission to deliberately reduce resolution per frame, thereby reducing bitrate.
Embodiments of the present invention are not focused on either of these uses, but rather find a third application for the super resolution technique: namely, to divide a given frame into a plurality of different lower resolution “projections” from which a higher resolution version of the frame can be reconstructed. Each projection is a version the same frame with a lower resolution than the original frame. The lower resolution samples of each different projection of the same frame have different spatial alignments relative to one another within the frame, so that the lower resolution samples of the different projections overlap but are not coincident. For example each projection is based on the same raster grid defining the size and shape of the lower resolution samples, but with the raster being applied with a different offset or “shift” in each of the different projections, the shift being a fraction of the lower resolution sample size in either the horizontal and/or vertical direction relative to the raster orientation.
An example is shown schematically in
A given frame F(t) comprises a plurality of higher resolution samples S′ defined by a higher resolution raster shown by the dotted grid lines in
The same frame F(t) is split into a plurality of different projections (a) to (d). Each of the projections of this same frame F(t) comprises a plurality of lower resolution samples S defined by applying a lower resolution raster to the frame, as illustrated by the solid lines overlaid on the higher resolution grid of in
Each lower resolution sample S represents a respective group of higher resolution samples S′ (each lower resolution sample covers a whole number of higher resolution samples). Preferably the value of the lower resolution sample S is determined by combining the values of the higher resolution samples, most preferably by taking an average such as a mean or weighted mean (although more complex relationships are not excluded). Alternatively the value of the lower resolution could be determined by taking the value of a representative one of the higher resolution samples, or averaging a representative subset of the higher resolution values.
The grid of lower resolution samples in the first projection (a) has a certain, first alignment within the frame F(t), i.e. in the plane of the frame. For reference this may be referred to here as a shift of (0, 0). The grids of lower resolution samples formed by each further projection (b) to (d) of the same frame F(t) is then shifted by a different respective amount in the plane of the frame. For each successive projection, the shift is by a fraction of the lower resolution sample size in the horizontal or vertical direction. In the example shown, in the second projection (b) the lower resolution grid is shifted right by half a (lower resolution) sample, i.e. a shift of (+½, 0) relative to the reference position (0, 0). In the third projection (c) the lower resolution grid is shifted down by another half a sample, i.e. a shift of (0, +½) relative to the second shift or a shift of (+½, +½) relative to the reference position. In the fourth projection the lower resolution grid is shifted left by another half a sample, i.e. a shift of (−½, 0) relative to the third projection or (0, +½) relative to the reference position. Together these shifts make up a shift pattern.
In
Note that the different projections do not necessarily need to be generated in any particular order, and any could be considered the “reference position”. Other ways of describing the same pattern may be equivalent. Other patterns are also possible, e.g. based on a lower resolution sample size of 4×4 higher resolution samples being shifted in a pattern of quarter sample shifts (a quarter of the lower resolution sample size).
The value of the lower resolution sample in each projection is taken by combining the values of the higher resolution samples covered by that lower resolution sample, i.e. by combining the values of the respective group of lower resolution samples which that higher resolution sample represents. This is done for each lower resolution sample of each projection based on the respective groups, thereby generating a plurality of different reduced-resolution versions of the same frame. The process is also repeated for multiple frames.
The effect is that each two dimensional frame now effectively becomes a three dimensional “slab” or cuboid, as shown schematically in
The projections of each frame are encoded and sent to a decoder in an encoded video signal, e.g. being transmitted over a packet-based network such as the Internet. Alternatively the encoded video signal may be stored for decoding later by a decoder.
At the decoder, each of the projections of the same frame can then be used reconstruct a higher resolution sample size from the overlapping regions of the lower resolution samples. For example, in the embodiment described in relation to
Each frame is preferably subdivided into a full set of projections, e.g. when the shift is half a sample each frame is represented in four projections, and in the case of a quarter shift into sixteen projections. Therefore overall, the frame including all its projections together may still represent the same resolution as if the super resolution technique was not applied.
However, unlike a conventional video coding scheme the frame is broken down into separate descriptions or sub-frames, which can be manipulated separately or differently. There are a number of uses for this, for example as follows.
Note also that, in embodiments, the multiple projections are created by a predetermined shift pattern, not signalled over the network from the encoder to the decoder and not included in the encoded bitstream. The order of the projection may determine the shift position in combination with the shift pattern.
An example communication system in which the various embodiments may be employed is described with reference to the schematic block diagram of
The communication system comprises a first, transmitting terminal 12 and a second, receiving terminal 22. For example, each terminal 12, 22 may comprise one of a mobile phone or smart phone, tablet, laptop computer, desktop computer, or other household appliance such as a television set, set-top box, stereo system, etc. The first and second terminals 12, 22 are each operatively coupled to a communication network 32 and the first, transmitting terminal 12 is thereby arranged to transmit signals which will be received by the second, receiving terminal 22. Of course the transmitting terminal 12 may also be capable of receiving signals from the receiving terminal 22 and vice versa, but for the purpose of discussion the transmission is described herein from the perspective of the first terminal 12 and the reception is described from the perspective of the second terminal 22. The communication network 32 may comprise for example a packet-based network such as a wide area internet and/or local area network, and/or a mobile cellular network.
The first terminal 12 comprises a tangible, computer-readable storage medium 14 such as a flash memory or other electronic memory, a magnetic storage device, and/or an optical storage device. The first terminal 12 also comprises a processing apparatus 16 in the form of a processor or CPU having one or more cores; a transceiver such as a wired or wireless modem having at least a transmitter 18; and a video camera 15 which may or may not be housed within the same casing as the rest of the terminal 12. The storage medium 14, video camera 15 and transmitter 18 are each operatively coupled to the processing apparatus 16, and the transmitter 18 is operatively coupled to the network 32 via a wired or wireless link. Similarly, the second terminal 22 comprises a tangible, computer-readable storage medium 24 such as an electronic, magnetic, and/or an optical storage device; and a processing apparatus 26 in the form of a CPU having one or more cores. The second terminal comprises a transceiver such as a wired or wireless modem having at least a receiver 28; and a screen 25 which may or may not be housed within the same casing as the rest of the terminal 22. The storage medium 24, screen 25 and receiver 28 of the second terminal are each operatively coupled to the respective processing apparatus 26, and the receiver 28 is operatively coupled to the network 32 via a wired or wireless link.
The storage medium 14 on the first terminal 12 stores at least a video encoder arranged to be executed on the processing apparatus 16. When executed the encoder receives a “raw” (unencoded) input video signal from the video camera 15, encodes the video signal so as to compress it into a lower bitrate stream, and outputs the encoded video for transmission via the transmitter 18 and communication network 32 to the receiver 28 of the second terminal 22. The storage medium on the second terminal 22 stores at least a video decoder arranged to be executed on its own processing apparatus 26. When executed the decoder receives the encoded video signal from the receiver 28 and decodes it for output to the screen 25. A generic term that may be used to refer to an encoder and/or decoder is a codec.
In operation, the projection generator 60 sub-divides each frame into a plurality of projections in the manner discussed above in relation to
In embodiments, each projection may be individually passed through the encoder 40 and treated as a separate stream. For encoding each projection may be divided into a plurality of blocks (each comprising a plurality of the lower resolution samples S).
Within a given projection, the forward transform module 42 transforms each block of lower resolution samples from a spatial domain representation into a transform domain representation, typically a frequency domain representation, so as to convert the samples of the block to a set of transform domain coefficients. Examples of such transforms include a Fourier transform, a discrete cosine transform (DCT) and a Karhunen-Loève transform (KLT) details of which will be familiar to a person skilled in the art. The transformed coefficients of each block are then passed through the forward quantization module 44 where they are quantized onto discrete quantization levels (coarser levels than used to represent the coefficient values initially). The transformed, quantized blocks are then encoded through the prediction coding stage 45 or 46 and then a lossless encoding stage such as an entropy encoder 48.
The effect of the entropy encoder 48 is that it requires fewer bits to encode smaller, frequently occurring values, so the aim of the preceding stages is to represent the video signal in terms of as many small values as possible.
The purpose of the quantizer 44 is that the quantized values will be smaller and therefore require fewer bits to encode. The purpose of the transform is that, in the transform domain, there tend to be more values that quantize to zero or to small values, thereby reducing the bitrate when encoded through the subsequent stages.
The encoder may be arranged to encode in either an inter prediction coding mode or an inter prediction coding mode (i.e. motion prediction). If using inter prediction, the inter prediction module 46 encodes the transformed, quantized coefficients from a block of one frame F(t) relative to a portion of a preceding frame F(t−1). The block is said to be predicted from the preceding frame. Thus the encoder only needs to transmit a difference between the predicted version of the block and the actual block, referred to in the art as the residual, and the motion vectors. Because the residual values tend to be smaller, they require fewer bits to encode when passed through the entropy encoder 48.
The location of the portion of the preceding frame is determined by a motion vector, which is determined by the motion prediction algorithm in the inter prediction module 46. According to embodiments of the present invention in which frames are each split into a plurality of projections, the motion prediction may be between two corresponding projections from different frames, i.e. between projections having the same shift within their respective frames. For example referring to
If using inter prediction, the transformed, quantized samples are subject instead to the intra prediction module 45. In this case the transformed, quantized coefficients from a block of the current frame F(t) are encoded relative to a block within the same frame, typically a neighbouring block. The encoder then only needs to transmit the residual difference between the predicted version of the block and the neighbouring block. Again, because the residual values tend to be smaller they require fewer bits to encode when passed through the entropy encoder 48.
In embodiments of the present invention, the intra prediction module 45 may have a special function of predicting between blocks from different projections of the same frame. That is, a block from one or more of the projections is encoded relative to a corresponding block in a base one of the projections. For example each lower resolution sample in one or more of the projections may be predicted from its counterpart sample in the base projection, e.g. so that the lower resolution sample S(m, n) in projection (b), (c) and (d) are each predicted from the sample S(m, n) in the first projection (a) and similarly for the other samples of each block. Thus the encoder only need to encode all but one of the projections in terms of a residual relative to the base projection.
This may present more opportunities for reducing the size of the residual, because corresponding counterpart samples from the different projections will tend to be similar and therefore result in a small residual. In embodiments the intra prediction module 45 may be configured to select which of the projections to use as the base projection and which to encode relative to the base projection. E.g. so the intra prediction module could instead choose projection (c) as the base projection and then encode projections (a), (b) and (d) relative to projection (c). The intra prediction module 45 may be configured to select which is the base projection in order to minimize or at least reduce the residual, e.g. by trying all or a subset of possibilities and selecting that which results in the smallest overall residual bitrate to encode.
Once encoded by the intra prediction coding module 45 or inter prediction coding module 46, the blocks of samples of the different projections are passed to the entropy encoder 48 where they are subject to a further, lossless encoding stage. The encoded video output by the entropy encoder 48 is then passed to the transmitter 18, which transmits the encoded video in one or more streams 33 to the receiver 28 of the receiving terminal 22 over the network 32, preferably a packet-based network such as the Internet.
In operation, each projection may be individually passed through the decoder 50 and treated as a separate stream.
The entropy decoder 58 performs a lossless decoding operation on each projection of the encoded video signal 33 in accordance with entropy coding techniques, and passes the resulting output to either the intra prediction decoding module 55 or the inter prediction decoding module 56 for further decoding, depending on whether intra prediction or inter prediction (motion prediction) was used in the encoding.
If inter prediction was used, the inter prediction module 56 uses the motion vector received in the encoded signal to predict a block from one frame based on a portion of a preceding frame. As discussed, this prediction could be between the same projection in different frames, or between different projections of different frames. In the latter case the motion vector and shift are added as shown in
If intra prediction was used, the intra prediction module 55 predicts a block from another block in the same frame. In embodiments, this comprises predicting blocks of one projection based on blocks of another, base projection. For example referring to
The decoded projections are then passed through the reverse quantization module 54 where the quantized levels are converted onto a de-quantized scale, and the reverse transform module 52 where the de-quantized coefficients are converted from the transform domain into lower resolution samples in the spatial domain. The dequantized, reverse transformed samples are supplied on to the super resolution module 70.
The super resolution module uses the lower resolution samples from the different projections of the same frame to “stich together” a higher resolution version of the frame. As discussed, this can be achieved by taking overlapping lower resolution samples from different projections of the same frame, and generating a higher resolution sample corresponding to the region of overlap. The value of the higher resolution sample is found by extrapolating between the values of the overlapping lower resolution samples, e.g. by talking an average. E.g. see the shaded region overlapped by four lower resolution samples S from the four different projections (a) to (d) in
In embodiments the process of reconstructing the frame from a plurality of projections may be lossless. For example this may be the case if each lower resolution sample represents four higher resolution samples of the original input frame as shown in
In other embodiments, the process may involve some degradation, and the higher resolution samples reconstructed at the decoder side need not be as high as the higher resolution sample size of the original input frame at the encoder side. For example this may be the case if each lower resolution sample represents four higher resolution samples of the original input frame, but only two projections are created e.g. with shifts of (0,0) and (+½, +½). In this case some information is lost in the process. However, the loss may be considered tolerable perceptually.
This process is performed for each a sequence of frames in the video signal being decoded. The reconstructed, higher resolution frames output for supply to the screen 25 so that the video is displayed to the user of the receiving terminal 22.
In one embodiment the different projections are transmitted over the network 32 from the transmitting terminal 12 to the receiving terminal 22 in separate packet streams. Thus each projection is transmitted in a separate set of packets making up the respective stream, preferably distinguished by a separate stream identifier for each stream included in the packets of that stream.
A result of transmitting in different streams is that one or more of the streams can be dropped, and it is still possible to decode at least a lower resolution version of the video from one of the projections, or potentially a higher (but not full) resolution version from a subset of remaining projections.
Projections may be dropped by the transmitting terminal 12 in response to feedback from the receiving terminal 22 or from the network 32 that there are insufficient resources at the receiving terminal or network conditions are inadequate to handle a full or higher resolution version of the video, or that a full or higher resolution is not required by the receiving terminal, or indeed if the transmitting terminal does not have enough resources to encode at a full or higher resolution. Alternatively or additionally, one or more of the streams carrying the different projections may be dropped by an intermediate element of the network 32 such as a router or intermediate server, in response to network conditions or information from the receiving terminal that there are insufficient resources to handle a full or higher resolution or that such resolution is not required.
For example, say a given frame is split into four projections (a) to (d) at the encoder side, each in a separate stream. If the receiving terminal 22 receives all four streams, the decoding system can recreate a full resolution version of that frame. If however one or more streams are dropped, e.g. the streams carrying projections (b) and (d), the decoding system can still reconstruct a higher (but not full) resolution version of the frame by extrapolating only between overlapping samples of the projections (a) and (c) from the remaining streams. Alternatively if only one stream remains, e.g. carrying projection (a), this can be used alone to display only a lower resolution version of the frame. Thus there may be provided a new form of layered or scaled coding based on splitting frames into different projections.
If prediction between projections is used then the base projection will not be dropped if it can be avoided, but one, some or all of the other projections predicted from the base projection may be dropped. To this end, the base projection is preferably marked as a priority by including a tag as side information in the encoded stream of the base projection. Elements of the networks 32 such as routers or servers may then be configured to read the tag (or note the absence of it) to determine which streams can be dropped and which should not be dropped if possible (i.e. dropping the higher priority base stream should be avoided).
In some embodiments a hierarchical prediction could be used, whereby one projecting is predicted from the base projection of the same frame, then one or more further projections are predicted in turn from each previously predicted projection of the same frame. E.g. so a second projection (b) may be predicted from a first projection (a), and a third projection (c) may be predicted from the second projection (b), and in turn a fourth projection (d) may be predicted from the projection (c). Further levels may be included if there are more than four projections. Each projection may be tagged with a respective priority corresponding to its order in the prediction hierarchy, and any dropping of projections or the streams carrying the projections may be performed in dependence on this hierarchical tag.
In embodiments the encoder uses a predetermined shift pattern that is assumed by both the encoder side and decoder side without having to be signalled between them, over the network, e.g. both being pre-programmed to use a pattern such as (0,0); (0, +½); (+½, +½); (+½, 0) as described above in relation to
Alternatively if the encoding system is configured to select which to use as a base projection, it may be that an indication concerning the shift pattern is included in the encoded signal. If any required indication is lost in transmission, the decoding system may be configured to use a default one of the projections alone so at least to be able to display a lower resolution version.
In further embodiments of the present invention the transform module 42 may be configured to exploit the different projections of the different frames in order to perform a three dimensional transform rather than two dimensional. As mentioned in relation to
As mentioned the purpose of performing a transform prior to quantization is that, in the transform domain, there tend to be more values that quantize to zero or to small values, thereby reducing the bitrate when encoded through the subsequent stages including the entropy encoding stage or the like. By arranging a frame into different offset projections and thereby enabling a three dimensional transform to be performed, there may be provided more instances where transformed coefficients quantize to zero or to smaller or more similar values for more efficient encoding by the entropy encoder 58.
A three dimensional transform explores redundancies between the coefficients of multiple two dimensional transformed regions that are created with multiple views. By selecting the views, as described herein, several representations or views of the same part of the frame can be generated. For natural images this preserves high local correlation between the pixels or samples. This high correlation is now presented in three dimensions instead of two and allows for more opportunities of quantizing transform coefficients which will result in more zero or small values.
It will be appreciated that the above embodiments have been described only by way of example.
For instance, the various embodiments are not limited to lower resolutions samples formed from 2×2 or 4×4 samples corresponding samples nor any particular number, nor to square or rectangular samples nor any particular shape of sample. The grid structure used to form the lower resolution samples is not limited to being a square or rectangular grid, and other forms of grid are possible. Nor need the grid structure define uniformly sized or shaped samples. As long as there is an overlap between two or more lower resolution samples from two or more different projections, a higher resolution sample can be found from an intersection of lower resolution samples.
The various embodiments can be implemented as an intrinsic part of an encoder or decoder, e.g. incorporated as an update to an H.264 or H.265 standard, or as a pre-processing and post-processing stage, e.g. as an add-on to an H.264 or H.265 standard. Further, the various embodiments are not limited to VoIP communications or communications over any particular kind of network, but could be used in any network capable of communicating digital data, or in a system for storing encoded data on a storage medium.
Other variants may be apparent to a person skilled in the art given the disclosure herein. The present various embodiments are not limited by the described examples but only by the accompanying claims.
