Method for assessing the quality of a distorted version of a frame sequence

This application claims the benefit, under 35 U.S.C. §119, of European Patent Application No. 08305139.1 of 30 Apr. 2008.

The invention is related to a method for assessing the quality of a distorted version of a frame sequence.

When image or frame sequences like video or film are compressed distortions are introduced in the compressed frame sequence. Errors introduced by transmission over a noisy or distorted channel may be another source of distortion.

For evaluating the quality of a distorted version of a frame sequence a video quality metric is necessary.

Well known mathematical metrics for measuring distances between an original and a distorted version, like mean square error or maximum absolute error, do not take into account the psychological aspects of distortions. That is, some distortions are more visible and thus more annoying to users than other distortions although the may all result in the same measured distance between distorted version and original.

Furthermore, the duration of a distortion plays a vital role with respect to the degree of annoyance the distortion introduces.

European Patent Application EP 1 804 519 describes a video quality objective assessment device for estimating subjective quality, said device includes a temporal/spatial feature amount derivation unit which derives a temporal/spatial feature amount which is the feature amount of deterioration.

There is a need for a method for assessing the quality of a distorted version of a frame sequence which reflects the influence of the human perceptive system better.

This is achieved by a method for assessing the quality of a distorted version of a frame sequence, said method comprising the steps of a) determining a last spatial distortion by comparing a block of a last frame of the sequence with a corresponding block of the distorted version of the last frame, b) determining, in a last-but-one frame, a best-matching block matching said block of the last frame best, c) determining a last-but-one spatial distortion by comparing the determined best-matching block of the last-but-one frame with a corresponding block of the distorted version of the last-but-one frame d) determining a spatio-temporal perceptual distortion value associated with said block using said determined distortions and using the determined spatio-temporal perceptual distortion value for assessing the quality.

By taking the temporal evolution of the spatial distortion associated with a block into account, the quality of the distorted video is assessed better.

The dependent claims describe advantageous embodiments of the method.

For instance, using said determined spatial distortions comprises using a first distortion gradient between the last and the last-but-one spatial distortion.

Or, the method further comprises iterating steps e) and f) through a number of frames preceding the last frame and using all determined spatial distortions for determining said spatio-temporal perceptual distortion value wherein using comprises using distortion gradients between consecutive frames.

If said first difference meets or exceeds a threshold, steps e) and f) are iterated through a first number of frames, otherwise, steps e) and f) are iterated through a second number of frames, said second number being smaller than said first number.

The method may further comprise counting a number of sign changes between the determined consecutive distortion gradients, determining a maximum among the determined distortion gradients and using said number of sign changes and said maximum for determining said spatio-temporal perceptual distortion value.

In an embodiment, the last frame (I 6 ) is partitioned into disjoint blocks and spatio-temporal perceptual distortion values are determined for each of said disjoint blocks. This allows for using the determined spatio-temporal perceptual distortions for determining a frame quality score for the last frame.

The method may even further comprise determining a frame quality score for each frame of the frame sequence, determining an average of the determined frame quality scores and using the determined average for assigning said quality.

Exemplary embodiments of the invention are illustrated in the drawings and are explained in more detail in the following description.

In the figures:

FIG. 1 shows an exemplary embodiment of the claimed invention

FIG. 2 shows an exemplary embodiment of the inventive spatio-temporal distortion evaluation module

FIG. 3 shows an exemplary embodiment of the inventive pooling module

FIG. 4 exemplarily depicts a spatio-temporal distortion tube and

FIG. 5 exemplarily depicts a convolution function used for temporal filtering of distortion gradients.

FIG. 1 shows an exemplary embodiment of the claimed invention. An original video OV is inputted into a motion estimation module MOES which outputs dominant motion parameters and motion vectors. A distorted version of the video DV is inputted into a spatial perceptual distortion evaluation module SPDE which outputs spatial perceptual distortion maps. Each spatial perceptual distortion map comprises distortion values associated with the pixels of one of the frames of the video. The spatial perceptual distortion maps, the motion parameters and the motion vectors are inputted into a spatio-temporal perceptual distortion evaluation module STPDE which outputs spatio-temporal perceptual distortion maps STPDM. Said spatio-temporal perceptual distortion maps STPDM are inputted into a pooling module POOL and may be used for other purposes, also.

The pooling module POOL then outputs video quality score VQS.

FIG. 2 shows an exemplary embodiment of the inventive spatio-temporal distortion evaluation module STPDE in more detail. The module STPDE comprises a sub-module STDTC which is responsible for creating so-called spatial distortion tubes associated with so-called block tubes TUBE. Said spatial distortion tubes are fed into two sub-modules TGC and TFSD. Sub-module TGC generates a so-called gradient tube from an inputted spatial distortion tube. The gradient tube is evaluated in sub-module MU whether it meets or exceeds a threshold. In dependency on the result one of at least two possible time parameter values which is provided to sub-module TFSD. Sub-module TFSD generates a so-called temporal filtered distortion tube from an inputted spatial distortion tube by temporarily filtering the inputted tube using the time parameter provided by module TGC.

The gradient tube generated by sub-module TGC is temporally filtered in sub-module TFDG. The temporally filtered gradient tube and the temporally filtered distortion tube are inputted into a sub-module for fixation pooling FIXP which outputs said spatio-temporal perceptual distortion maps STPDM.

FIG. 3 shows an exemplary embodiment of the inventive pooling module POOL in more detail. Said pooling module POOL comprises a sub-module for spatial pooling SPOOL and a sub-module for temporal pooling TPOOL. The spatial pooling SPOOL generates a frame quality score for each frame from the spatio-temporal perceptual distortion maps which it receives as input. Said frame quality scores are combined into a video quality score VQS in sub-module TPOOL. Thus, the pooling module POOL performs the steps of calculating quality scores for the individual frames as well as for the whole video sequence from the spatio-temporal perceptual distortion maps.

The steps performed by the modules and sub-modules are explained in more detail, below.

The step performed in the Spatial Perceptual Distortion Evaluation module SPDE is the computation of spatial perceptual distortion maps VE t,x,y . For each frame t of the video sequence, a spatial perceptual distortion map is computed. Each entry (x,y) of this spatial distortion maps represents the degree of distortion that is perceived between a pixel at position (x,y) in the original frame and a distorted version of the pixel at the same position in the distorted version of the frame. In an exemplary embodiment, the degree of distortion is determined by help of a distortion measure described in a previous invention filed as European Patent Application, filing No. 08300006.7, which is enclosed entirely into this application by reference. Others well-known metric may be used in addition or alternatively. Such other metrics are, for instance, Mean Square Error (MSE), Peak-Signal-to-Noise-Ratio (PSNR) or Structural SIMilarity (SSIM).

Prior to, in parallel to or after calculating the spatial perceptual distortion maps VE t,k,l , the step of motion estimation in the original video frame sequence is performed. For the blocks of a frame, corresponding best matching blocks in an immediately preceding frame and corresponding replacement or motion vectors are determined. From said motion vectors, dominant motion parameters are calculated. Dominant motion parameters may be determined by different means such as, for instance, Least-Mean Square or algorithm belonging to the family of the M-estimator.

These steps (dominant motion and motion vectors estimation) may be achieved with a Hierarchical motion estimator (HME).

The steps performed in modules SPDE and MOES are necessary for obtaining the spatial distortion maps and the motion information required for the inventive spatio-temporal perceptual distortion evaluation.

The step performed in the Spatio-temporal Perceptual Distortion Evaluation module STPDE is calculation of spatio-temporal perceptual distortion maps VE t,k,l from the spatial distortion maps and the motion information. The spatio-temporal perceptual distortion maps VE t,k,l comprise spatio-temporal perceptual distortion values associated with blocks. This step comprises determining temporally filtered spatial distortion maps VE t,k,l by filtering the spatial distortions determined for blocks of a frame with spatial distortions determined for corresponding predecessor blocks in a preceding frame, said predecessor blocks being determined by help of the motion information. E.g., for each block a weighted sum of the block's spatial distortion and of the corresponding predecessor block's spatial distortion is determined by said filtering. Furthermore, temporal distortion gradient maps GR t,k,l are determined. Each entry (k,l) of these temporal gradient distortion maps GR t,k,l represents the difference in the degree of distortion determined for a block at position (k,l) in the frame t and the degree of distortion determined for a predecessor block of said block at position (k,l), said predecessor block being in the preceding frame.

From the temporally filtered spatial distortion maps VE t,k,l and the temporal gradient maps GR t,k,l , spatio-temporal perceptual distortion maps VE t,k,l are computed.

The time horizon used for temporal filtering may be chosen according to computational or other requirements. That is, predecessors of said predecessor blocks may be determined by help of the motion information and the spatial distortions determined for said predecessors of predecessor blocks may be used for filtering, too. More generally, spatial distortions determined for antecedent blocks from all frames that precede a given frame within the time horizon may be used for temporally filtering spatial distortions determined for blocks of the given frame. Exemplarily, the time horizon is set to 400 ms which corresponds to the time scale of the human eye fixation.

This concept is exemplarily depicted in FIG. 4 . The figure depicts an example series or sequence of images or frames I 1 . . . I 6 . A path or history TUBE, also called a block tube, of an arbitrarily chosen block B 6 of the last frame I 6 is depicted as a rectangular tube. Only the block tube TUBE of arbitrarily chosen block B 6 is depicted but corresponding block tubes are formed for all the disjoint blocks of the last frame I 6 . The block tube TUBE associated with arbitrarily chosen block B 6 is determined by a series of best-matching blocks B 5 , B 4 , B 3 , B 2 , B 1 wherein a first best-matching block B 5 in last-but-one frame I 5 matches said arbitrarily chosen block B 6 in last frame I 6 best. A second best-matching block B 4 in last-but-two frame I 4 matches said best-matching block of last frame I 6 best, and so on, up to a fifth best-matching block B 1 in last-but-five frame I 1 matching a fourth best-matching block B 2 of last-but-four frame I 2 best. Said fourth best-matching block B 2 in last-but-four frame I 2 matches a third best-matching block B 3 of last-but-tree frame I 3 best which matches said second best-matching block B 4 of last-but-two frame I 4 best.

Furthermore, the time horizon may be shortened if the object inside the block appears or disappears in the scene. A method to detect the appearance or the disappearance of an object rests on the use of the dominant motion. Indeed, each block is classified by comparing is motion with a parametric dominant motion: it means that each block is either inlier or outlier to the dominant motion. A classification of inlier-outlier-change of a block between two consecutive frames means the appearance or the disappearance of the object it belongs to, and so indicates the limit of the temporal horizon of this block.

Corresponding to the block tube TUBE there are a spatial distortion tube VE t,k,l tube and a distortion gradient tube GR t,k,l tube . The spatial distortion tube VE t,k,l tube is a sequence of the spatial distortion values associated with the blocks in the block tube TUBE. And the distortion gradient tube GR t,k,l tube is a sequence of the differences between consecutive spatial distortions in the spatial distortion tube. Therefore, the distortion gradient tube GR t,k,l tube comprises one element less than the spatial distortion tube or the block tube TUBE.

In a block tube TUBE, the distortion gradient at time instant t i is computed as follows:

Gr t i ,k,l tube =VE t i ,k,l tube −VE t i−1 ,k,l tube   (1)

wherein VE t i ,k,l tube is the distortion value at instant t i , VE t i−1 ,k,l tube is the distortion value at instant t i −1 and t i is comprised in the time horizon.

The spatial distortion tube associated with arbitrarily chosen block B 6 is used for filtering the spatial distortion associated with said block B 6 . That is, spatial distortions determined for said best-matching blocks are used for temporally filtering the spatial distortion of the arbitrarily chosen block B 6 . Said filtering may be recursive, that is, temporally filtered spatial distortions determined for said best-matching blocks are used for temporally filtering the spatial distortion of the arbitrarily chosen block B 6 .

The temporal filtering of spatial distortions may further depend on a time parameter adjusting the influence of the spatial distortions or the temporally filtered spatial distortions of predecessors on the filtering in dependency on their distance in time to a frame under inspection. That is, with a larger time parameter the filtering distortions of all predecessors in the time horizon are influential for filtering while, with a smaller time parameter, filtering is dominated or mainly guided by distortions of predecessors in narrow temporal vicinity of the last frame under inspection.

The time parameter used for filtering arbitrarily chosen block B 6 may be adjusted in dependency on the associated distortion gradient tube. If the difference between the spatial distortions determined for arbitrarily chosen block B 6 and a block of an immediately preceding frame best-matching said arbitrarily chosen block B 6 meets or exceeds a threshold, a first time parameter, for instance 400 ms, is used for temporal filtering. Otherwise, a smaller second time parameter, for instance 200 ms, is used for filtering. In this case, the distortion gradient associated with arbitrarily chosen block B 6 may also be set to Zero.

The distortion gradient associated with arbitrarily chosen block B 6 may also being subjected to temporal filtering by help of the distortion gradient tube. Said filtering of the distortion gradient is based on a maximum distortion gradient MaxGr t,k,l tube in the distortion gradient tube and the number of sign changes NS t,k,l tube between consecutive distortion gradients in the distortion gradient tube. The number of sign changes is used to weight the maximum distortion gradient. For smoothing, the number of sign changes may be convoluted, for instance with the following Gaussian convolution function which is further exemplarily depicted in FIG. 5 :

The function fs( ) gives more importance to temporal distortion at certain sign change frequencies. Good results are achieved with g s =16, σ s =6 and μ s =1.3 but other parameter settings are possible, too. The result of this step is the map VÊ t,k,l tube where each block (k,l) is the result of the temporal filtering of the distortion gradient in each tube beginning at the frame t.

Then, the temporally filtered distortion gradient associated with a block at position (k,l) in frame t may be obtained following this equation:

VÊ t,k,l tube =Max Gr t,k,l tube ·fs ( NS t,k,l tube )  (3)

Sub-module FIXP performs a fixation pooling, in which the map VE t,k,l tube and the map VÊ t,k,l tube are combined in order to obtain the final spatio-temporal distortion maps VE t,k,l . These maps may be computed following this equation:

VE t,k,l = VE t,k,l tube ·(α+β· VÊ t,k,l tube )  (4)

Parameters α and β are empirically deduced. For instance, good results where achieved with α being 1 and β being 3.

Module POOL is composed of 2 sub-modules, namely a spatial pooling sub-module SPOOL and a temporary pooling sub-module TPOOL as depicted in FIG. 3 . Said pooling sub-modules perform the following steps:

First, a frame quality score Q t is computed from the spatio-temporal distortion maps using a Minkowski summation:

wherein K and L are the height and the width of the spatio-temporal distortion maps respectively, i.e. the vertical and the horizontal number of blocks in the frame at time t, and β p is the Minkowski exponent.

Then, the final video quality score VQS is computed from every frame quality scores Q t of the video. The final VQS is computed as follows:

VQS= Q +min( q 1 ·max(Δ t ), q 2 · Q )  (6)

wherein Q is the average of Q t over all frames of the video, and max(Δ t ) is a term representing the maximum among the variations of quality along the sequence wherein Δ t is:

wherein function abs( ) returns the absolute value of its argument.

Good results where achieved with m=2, p=1, q 1 =1 . and q 2 =0.5 but other parameter settings are possible.

The objective quality score VQS is possibly transformed into predicted Mean Observer Score (MOSp) using a psychometric function. The psychometric function is given by:

wherein b 1 , b 2 and b 3 are the parameters of the psychometric function which are determined by comparing mean observer scores assigned to videos in a database and objective quality scores VQS calculated for the same videos.

The range of the objective quality score VQS is between [0; 1[, and the range of the MOSp is between [0; 6[.

The corresponding category rating is the following:

5 is imperceptible; 4 is perceptible but not annoying; 3 is slightly annoying; 2 is annoying; 1 is very annoying.

The invention may be used advantageously for assessing video quality for instance in an encoding device to compare a reconstructed video with the corresponding video source.

The invention may further be used advantageously for comparing the performance of encoding devices using different codec.

The perceptual error map computed according to the invention may also be used in an encoding device to locally adapt the decision or the quantification step.