JP2014068064A - Objective evaluation device of video quality - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an objective evaluation device of video quality which can obtain an objective evaluation value by reducing the influence of intrinsic spacial and temporal properties of an evaluated video.SOLUTION: A motion detection unit 1 determines a motion vector in an evaluated frame with reference to a preceding frame in a frame memory 2, a prediction signal generation unit 4 generates a prediction signal by using the motion vector, and a residual signal calculation unit 5 generates a residual signal by taking the difference of the prediction signal and a pixel signal. A block distortion feature amount calculation unit 6 calculates a block distortion feature amount by using a difference absolute value in the block boundary out of the residual signals. A flicker feature amount calculation unit 7 calculates a flicker feature amount by using the inter-frame difference power of the residual signal. A feature amount integration unit 8 determines an objective evaluation value on the basis of the block distortion feature amount and flicker feature amount.

Description

  The present invention relates to an objective evaluation apparatus for video quality, and in particular, the perceptual quality of an image deteriorated due to compression encoding of a moving image and a reproduction failure caused by a transmission path failure, the baseband signal of a decoded image without using a reference image. The present invention relates to an objective evaluation apparatus for video quality estimated from analysis alone.

  When storing / transmitting digital video, the amount of information is usually reduced by compression encoding. This compression encoding generally means lossy compression, and the decoded image is controlled to minimize visual degradation for a given compression rate (bit rate), but it is On the other hand, it has a feature of having an error corresponding to the image feature.

  Here, the image features refer to features such as the fineness, the magnitude of movement, and the complexity of the components in the screen. As described above, visual degradation due to lossy compression is designed to be minimal. In general, lossy encoding varies the degree of degradation recognized by the above-described image features even if the compression rate is constant. It has the property. For this reason, there is a need for a technique for quantitatively measuring image quality degradation accompanying lossy compression.

  Conventional measurement of image quality has been performed by subjective evaluation based on human eyes. Subjective evaluation requires knowledge and experience to correctly understand and implement a given evaluation method and the time and cost burden of gathering a large number of non-experts and presenting videos to individual subjects. The implementation is not easy. Accordingly, it is desired to establish an objective image quality evaluation method that extracts video features (features) quantified by analysis of video signals and derives the quality of the video from the video feature values.

  The objective image quality evaluation method uses reference information (reference; original image before compression coding, etc.) in addition to the video signal to be evaluated, or whether the video signal to be analyzed is a compressed signal or an uncompressed video signal. It is classified into multiple frameworks.

  In the former point of view, there is a classification of FR (Full Reference) / RR (Reduced Reference) using reference information and NR (No Reference) not using reference information. In the latter point of view, there is a bit stream type / There is a classification of a base type that analyzes an uncompressed signal / a hybrid type that uses a bit stream and a base band together.

  In general, in objective image quality evaluation, the accuracy of subjective image quality estimation is higher when there is reference information serving as a quality standard, but reference information is not always obtained depending on the embodiment (application) (eg, video transmission). Assuming monitoring on the receiving side of the system, the image on the transmitting side cannot be referred to), and there is a problem that the application range is narrow. Also, referring to the compressed bitstream may require descrambling, demultiplexing, and the like, and the configuration that can be implemented is limited compared to referring only to the baseband signal that the video receiver always has as an output. Sometimes.

  As described above, the baseband NR objective evaluation method that estimates the subjective image quality by analyzing only the pixel information contained in the baseband signal without reference information is available for various embodiments, although there is a problem with the subjective image quality estimation accuracy. This is desirable from the viewpoint of

  As a parametric NR type evaluation method, there is Patent Document 1 relating to the application of the present inventors.

  Further, as a baseband NR type objective evaluation method, there is Patent Document 2 relating to the application of the present inventors. The invention disclosed in Patent Document 2 is based on the assumption that subjective quality degradation due to coding degradation is estimated, and the video feature amount is analyzed by baseband signal analysis based on block distortion and flicker intensity that are easily recognized visually as coding degradation. The subjective image quality is estimated by integrating the two feature quantities.

JP2011-199380 (Image quality objective image quality evaluation device and automatic monitoring device) JP 2010-124104 (Objective Image Quality Evaluation System)

  In the invention disclosed in Patent Document 2, a Hadamard transform coefficient component (a base image is rectangular and its boundary coincides with an encoded block boundary) in the calculation of the block distortion feature amount in the baseband signal of the decoded image. The difference power between frames is used to calculate the flicker feature amount. These are defined in consideration of the characteristics of block distortion / flicker error signals, but if the target image originally has a rectangular pattern or has a fast movement, it is caused by coding deterioration. In some cases, it is difficult to determine whether the video to be evaluated is due to the spatiotemporal nature of the video to be evaluated. If the two cannot be correctly identified, an objective evaluation is performed with an unduly high estimation of the strength of encoding degradation, and the subjective image quality estimation accuracy based on the evaluation is greatly impaired.

  In view of the above-described problems, an object of the present invention is to provide an objective video quality evaluation apparatus capable of obtaining objective evaluation values with higher accuracy by reducing the influence of spatiotemporal properties inherent in video to be evaluated. There is.

  In order to solve the above problems, the present invention provides a video quality objective evaluation apparatus that evaluates image quality using only pixel signal information of an evaluation image without using a reference image. A motion detection unit that detects a motion vector from each pixel signal of the frame, a prediction signal generation unit that generates a prediction signal from the motion vector and the pixel signal of the evaluation image, the prediction signal, and a pixel signal of the evaluation image target frame Based on a residual signal calculation unit that generates a residual signal as a difference, a block distortion feature amount calculation unit that calculates a block distortion feature amount in each frame from the residual signal, and an average in the video sequence of the block distortion feature amount And a feature amount integration unit for deriving an objective evaluation value of video image quality.

  The present invention further includes a flicker feature amount calculation unit that calculates a flicker feature amount in each frame from the residual signal, and the feature amount integration unit includes a video sequence of each of the block distortion feature amount and the flicker feature amount. A second feature is to derive an objective evaluation value of video quality based on the inner average.

  According to the first feature, since the block distortion feature value is calculated from the residual signal, the objective evaluation value is estimated with high accuracy by reducing the influence of the spatial property inherent in the video to be evaluated. Can do.

  According to the second feature, since the flicker feature is calculated from the residual signal, the objective evaluation value can be estimated with high accuracy by reducing the influence of the spatial property inherent in the video to be evaluated. Can do.

It is a functional block diagram of the objective evaluation apparatus for video quality according to an embodiment. It is a figure which shows the example for demonstrating the definition of a block distortion feature-value.

  FIG. 1 is a functional block diagram of an objective video quality evaluation apparatus according to an embodiment. The objective evaluation device 10 for video quality includes a motion detection unit 1, a frame memory 2, a global motion vector (GMV) selection unit 3, a prediction signal generation unit 4, a residual signal calculation unit 5, a block distortion feature amount calculation unit 6, a flicker A feature amount calculation unit 7, a feature amount integration unit 8, a subjective evaluation value estimation unit 9, and a switch 11 are provided.The uncompressed video signal is input, the objective evaluation value is one output, and further calculation is performed using the output. By doing so, the estimated value of the subjective evaluation value is used as another output. Below, each said function part is demonstrated.

[About input video signal]
Uncompressed video signals are input to the video quality objective evaluation apparatus 10 in the order of frame numbers, and the input video signals are passed to the motion detection unit 1, the frame memory 2, and the residual signal calculation unit 5.

[About frame memory 2]
The frame memory 2 holds the input video signal in units of frames and provides it to the motion detection unit 1, so that when the motion detection unit 1 detects a motion vector between adjacent frames, the frame memory 2 It is possible to refer to the information of the frame.

[About motion detector 1]
The motion detection unit 1 divides a frame obtained from the input video signal into small pixel blocks, applies motion prediction to the divided blocks with reference to the previous frame held in the frame memory 2, and generates a motion vector ( MV) is detected. The MV detection accuracy may be integer accuracy, and it is not necessary to detect the motion of the decimal pixel unit using motion compensation as in the case of moving image coding.

Any method of MV detection may be selected. As an example, MV; v _B = (v _B0 , v _B1 ) by a simple block matching method is defined as (Equation 1) below.

Here, x _{f, B} (n ₀ , n ₁ ) represents the luminance signal (pixel signal) value at the frame position f, the pixel position (n ₀ , n ₁ ) in the pixel block B, and the square of the luminance signal difference MV is detected as the one that minimizes the sum. It is desirable that the size of the pixel block matches the control unit in the assumed encoding method. For example, in the case of H.264, 16 × 16 pixels or an integer multiple thereof is preferable as in the macroblock (MB) size. If the encoding scheme applied to the video to be evaluated is known, the predetermined size of the corresponding control unit may be used. If the encoding scheme is unknown, the predetermined size assumed in some encoding scheme may be used by user designation or the like. That's fine.

[About switch 11]
The switch 11 is set to be connected to either the contact A or B, and when connected to the contact A, the first embodiment uses the GMV selection unit 3, and when connected to the contact B, the second skips the GMV selection unit 3. This corresponds to the embodiment. Note that the switch 11 is fixedly set by user designation or the like, and the connection state of the contact A or B is not switched in the middle of the evaluation target video.

  In the first embodiment, the GMV selection unit 3 derives a GMV from the motion vector detected by the motion detection unit 1 and inputs the GMV to the prediction signal generation unit 4. In the second embodiment, the motion vector detected by the motion detector 1 is input to the prediction signal generator 4 as it is. In any embodiment, the motion vector detected by the motion detector 1 is also input to the flicker feature quantity calculator 7.

[About GMV Selector 3]
In the present invention, in the case of the second embodiment, the MV of each pixel block output from the motion detection unit 1 in the prediction signal generation unit 4 is used in order to separate the characteristic originally possessed by the input video and the deteriorated signal component. The prediction image is generated by using this, and the residual signal calculation unit 5 obtains the residual image.

  However, if the MV is used directly, if the movement is not uniform and the MVs do not match between adjacent blocks (eg lawn, water surface, etc.), it is not caused by the degradation signal component that is being detected. Due to the MV selection variation of the motion detection unit 1, a large edge (block distortion) due to signal discontinuity may be detected at the pixel block boundary of the residual image.

  Therefore, in the first embodiment, the GMV selection unit 3 derives a global motion vector (Global MV; GMV) from the MV of each pixel block output from the motion detection unit 1, thereby matching the MVs between adjacent blocks as much as possible. Like that.

Any method may be used as a method for deriving GMV. As an example, in Dirac encoding, a vector G = Av + b (A and b are affine transformation matrices and offsets) obtained by performing affine transformation on a local vector v (MV detection result of the motion detection unit 1) is GMV, It is derived by finding an affine transformation matrix that minimizes the L2 norm || G-v || ² of the difference between the local vector and GMV.

[About predicted signal generator 4]
The prediction signal generation unit 4 directly uses the motion vector (local vector) detected by the motion detection unit 1 in the second embodiment, and uses the GMV derived by the GMV selection unit 3 in the first embodiment ( The prediction signal x _{pred, B} (n ₀ , n ₁ ) is generated by Equation 2) and input to the residual signal calculation unit 5. Note that v _B = (v _B0 , v _B1 ) is a local vector or GMV.

[Residual signal calculator 5]
The residual signal calculation unit 5 obtains the difference at the same pixel position (n ₀ , n ₁ ) between the input video signal and the prediction signal as shown in (Equation 3), and obtains the residual signal r _{f, B} (n ₀ , n ₁ ) and the residual signal is input to the block distortion feature value calculation unit 6 and the flicker feature value calculation unit 7.

[Block distortion feature quantity calculation unit 6]
The block distortion feature quantity calculation unit 6 outputs a block distortion feature quantity for each frame as an index indicating the degree of visual recognition of block distortion, which is spatial degradation on the decoded image, and inputs it to the feature quantity integration unit 8. At this time, the block distortion feature value for each block is first obtained as described below, and the block distortion feature value for each frame is finally output as an average of those satisfying a predetermined condition in consideration of visual characteristics.

  When the quantization step differs greatly between orthogonal transform blocks, block distortion is such that the prediction residual signal becomes discontinuous at the boundary of motion prediction unit blocks (in many coding systems, this is a macroblock). This is considered to cause visual detection on the luminance signal.

  Furthermore, it is considered that a discontinuous component on the luminance signal due to a spatial characteristic originally possessed by the image instead of block distortion does not cause a large discontinuity on the prediction residual signal. Therefore, in the present invention, the block distortion feature amount is defined based on the discontinuity of the prediction residual signal at the boundary portion of the unit block for motion prediction.

FIG. 2 is a diagram showing an example for explaining the definition of the block distortion feature quantity. Here, the unit block for motion detection (second pixel block) is 2N _M pixels × 2N _M lines, and the unit block for orthogonal transformation Although (the first pixel block) is an example where a N _M pixels × N _M line, the relationship of magnitude between the unit block can be obtained similarly block distortion characteristic amount may be different.

In FIG. 2, the unit blocks B ₀ , B ₁ , B _{2 for} orthogonal transformation belong to different motion detection blocks A, B, C, respectively, and a large luminance change occurs at the boundary between these motion detection blocks It is assumed that block distortion is detected by Here, when the block distortion feature amount of the block B ₀ is obtained, it follows a boundary portion (indicated by a bold line in the figure) between _two blocks B ₁ and B ₂ that belong to different and different motion detection blocks. The absolute difference value of the prediction residual signal is obtained between pixels (indicated by a bidirectional arrow in the figure). Block distortion characteristic quantity SI in the block B ₀ (B ₀₎ can be expressed as follows (Equation 4).

That is, among the 2 × 2 orthogonal transform unit blocks present in the motion detection block A, in the block B ₀ which is the lower right part, the right end {(N _M −1, j) | j ∈ border} There is a boundary with block B ₁ in another motion detection block B, and there is a boundary with block B ₂ in another motion detection block C at the lower end {(i, N _M −1) | i ∈ border} Therefore, in (Equation 4), the sum of the absolute difference values of the prediction residual signals at these boundaries is taken. For the pixels (N _M −1, N _M −1) at the right end and the lower end of the block B ₀ , absolute values of differences from the respective pixels of the blocks B ₁ and B ₂ are obtained.

Similarly, if the feature amount SI (B ₁ ) of the block B ₁ is to be obtained, the upper end portion and the right end portion thereof are used as boundaries, and the feature amount SI (B ₂ ) of the block B ₂ is to be obtained. The sum of absolute difference values may be obtained at each boundary portion using the portion and the lower end portion as the boundary portion.

  Note that, as (Equation 4), the embodiment (2) is different from the embodiment (1) in which the block distortion feature amount is obtained by limiting the boundary between different motion detection unit blocks (second pixel blocks). There may be no limitation between motion detection unit blocks. In this case, regardless of the motion detection unit block to which it belongs, the sum of the absolute difference values of the prediction residual signal is obtained in the entire boundary portion of the orthogonal transform unit block (first pixel block).

In the embodiment (2), for example, in the case of the block B ₀ , the same calculation is performed using not only the right end portion and the lower end portion but also the left end portion and the upper end portion as the boundary portions as in the first embodiment. However, the embodiment (1) is preferable in the present invention. In addition, if the motion detection unit block and the orthogonal transform unit block match, the embodiment (1) and the embodiment (2) match.

  The block distortion feature value calculation unit 6 finally obtains the block distortion feature value for the frame as an average of the block distortion feature values for each block. However, it is known that when the signal variance in the pixel block is large, the luminance change at the block boundary is hardly perceived as block distortion in terms of visual characteristics. Therefore, the average of the spatial degradation feature amounts of blocks whose signal variance is equal to or less than a given threshold value is obtained.

Here, the number of blocks signal distributing var (B), a given threshold sigma _TH, N _V the signal variance var (B) is equal to or less than a given threshold value sigma _TH of the luminance component (pixel signals) of the block B Then, the block distortion feature amount SI [f] of the frame f is expressed as (Equation 5) below.

Note that the block B is an orthogonal transform block (B ₀ , B ₁ , B ₂ , etc. in the example of FIG. 2) that is a unit of calculation of the block distortion feature value SI (B). Absent. Further, the signal variance var (B) is the variance of the pixel signal, not the variance of the prediction residual signal.

  Note that the size of the unit block for orthogonal transform may be set in the same manner as described for setting the size of the unit block for motion detection in the description of the motion detection unit 1. That is, if the size used in the encoding method is known, the known size may be set, and if it is unknown, the predetermined size assumed in some encoding method may be set.

[Flicker feature calculation unit 7]
The flicker feature amount calculation unit 7 first obtains a flicker feature amount representing the degree of influence of each frame on the flicker degradation of the video as an average of the difference power between frames of the residual signal in the target frame and the immediately preceding frame, and further calculates the entire video. The flicker feature amount is obtained and input to the feature amount integration unit 8. At this time, using the MV input from the motion detection unit 1, the average power is obtained in a part that can actually affect the degradation when visual characteristics are taken into consideration.

  Flicker degradation is detected when there is a large quality fluctuation at every intra-frame insertion period of motion compensation predictive coding, and is perceived by a sudden change in luminance between successive frames. . In the present invention, it is considered that most of the motion inherent in the image to be evaluated has been removed by the derivation of the residual signal in the residual signal calculation unit 5. Therefore, if there is a characteristic that the intensity of the residual signal changes greatly in time, it is not caused by the temporal characteristic (that is, the motion characteristic) that the image originally had, but is accompanied by compression coding. This is thought to be due to deterioration.

However, it is expected that masking is visually masked even in the case of a large luminance change in an area where the movement is large. Therefore, as shown in the following (Equation 6), the flicker feature amount calculation unit 7 performs the residual signal interframe difference in the block only when the size of the local vector of the motion detection block is equal to or less than a given threshold V _TH. By calculating the average power, the flicker feature amount TI [f] of the frame f is obtained.

Here, as described above, the local vector is the MV itself derived by the motion detection unit 1, and is particularly called a local vector in order to particularly distinguish it from the GMV by the GMV selection unit 3. In (Equation 6), _NT is the number of blocks whose local vector size is less than a given threshold.

[About the feature integration unit 8]
The feature amount integration unit 8 uses the block distortion feature amount SI [f] and the flicker feature amount TI [f] defined for each frame f and derived by the block distortion feature amount calculation unit 6 and the flicker feature amount calculation unit 7, respectively. Thus, an objective evaluation value is derived.

For this reason, by first obtaining the average of the SI [f] and TI [f], respectively, the block distortion feature SI of the entire sequence (f ∈ sequence) as shown in the following (Equation 7) and (Equation 8): And the flicker feature quantity TI is obtained. In the following equations, N _F is the total number of frames in the sequence.

Then, the feature quantity integration unit 8 obtains the objective evaluation value Q _obj by the following (formula 9). Here, c0, c1, and c2 are constants for weighted sums, and α and β are constants as power exponents for further modulating the weighted sums.

[Subjective evaluation value estimation unit 9]
The subjective evaluation value estimation unit 9 obtains an estimated value Q _est of the subjective evaluation value by applying the conversion function f to the derived objective evaluation value Q _obj as shown in the following (Equation 10).

The above-described conversion function has a characteristic that takes into consideration a non-linear characteristic in perceptual psychology, and a non-linear function such as a monotonic cubic function or a logistic function is preferable. The monotonic cubic function has the function form shown in (Formula 11) below. The logistic function has the function form shown in (Formula 12) below. In (Expression 11) and (Expression 12), k _i (i = 0,1,2,3) is a constant.

The constants c _i (i = 0,1,2), α, β, k _i (i = 0,1,2,3) in the above equations (Equation 9) to (Equation 12) are the subjective evaluation values. It is set in advance so that the correlation between the estimated value Q _est and the true subjective evaluation value is maximized. Specifically, using a plurality of evaluation images with known subjective evaluation values, the sum of squared differences between the estimated value Q _est obtained based on the above definition and the true subjective evaluation value (measured value obtained by a subjective evaluation experiment) The coefficient that minimizes is obtained by multiple regression analysis.

[Other Embodiments and Supplementary Items of the Present Invention]
As an alternative embodiment to the configuration of FIG. 1, the video quality objective evaluation device 10 is configured to include only one of the block distortion feature amount calculation unit 6 and the flicker feature amount calculation unit 7 In this case, the objective evaluation value and the estimated value of the subjective evaluation value are given based on only one of the block distortion feature value and the flicker feature value.

  In the other embodiment, the feature amount integration unit 8 receives only one of the block distortion feature amount or the flicker feature amount as an input and considers that the other feature amount not received is zero or (Equation 9) By assuming that there are no terms for the block distortion feature value SI and the flicker feature value TI, it is possible to function in the same manner as in the embodiment using both of the above.

  As a supplementary matter in the block distortion feature quantity calculation unit 6 and the flicker feature quantity calculation unit 7, in these functional units, as shown in (Equation 5) and (Equation 6), both threshold conditions (threshold conditions relating to pixel signal dispersion, respectively) And a threshold condition regarding the size of the local vector) to limit the target. The threshold condition is preferably used, but may not be used.

  That is, instead of (Equation 5), the block distortion feature quantity SI [f] of the frame f may be obtained as an average of all the block distortion feature quantities of the block B belonging to the frame f regardless of the variance value. Instead of 6), the flicker feature quantity TI [f] of the frame f may be obtained by obtaining the residual power inter-frame difference power for all locations in the frame f regardless of the size of the local vector.

  10 ... Objective evaluation device for video quality, 1 ... motion detection unit, 2 ... frame memory, 3 ... GMV (global motion vector) selection unit, 4 ... predicted signal generation unit, 5 ... residual signal calculation unit, 6 ... block distortion Feature amount calculation unit, 7 ... Flicker feature amount calculation unit, 8 ... Feature amount integration unit, 9 ... Subjective evaluation value estimation unit

Claims (10)

In a video quality objective evaluation apparatus that evaluates image quality using only pixel signal information of an evaluation image without using a reference image,
A motion detector that detects a motion vector from each pixel signal of the evaluation target frame in the evaluation image and the immediately preceding frame;
A prediction signal generation unit that generates a prediction signal from the motion vector and the pixel signal of the evaluation image; and a residual signal calculation unit that generates a residual signal as a difference between the prediction signal and the pixel signal of the evaluation image target frame;
A block distortion feature amount calculation unit for calculating a block distortion feature amount in each frame from the residual signal;
A video quality objective evaluation apparatus, comprising: a feature value integration unit that derives an objective evaluation value of video quality based on an average of the block distortion feature values in a video sequence.   2. The video quality according to claim 1, wherein the block distortion feature value calculation unit calculates the block distortion feature value based on an absolute difference value of the residual signal at a boundary of a predetermined first pixel block. Objective evaluation device. The motion detection unit detects the motion vector in units of a predetermined second pixel block,
The block distortion feature amount calculation unit is configured to calculate the block distortion based on a difference absolute value of the residual signal only in a portion of the boundary of the predetermined first pixel block that matches the boundary of the predetermined second pixel block. The objective evaluation apparatus for video quality according to claim 2, wherein the feature quantity is calculated.   The block distortion feature value calculation unit calculates the block distortion feature value based on the absolute value of the difference of the residual signal only at a location where a variance of pixel signals in the predetermined first pixel block satisfies a predetermined threshold condition. 4. The video quality objective evaluation device according to claim 2, wherein the objective video quality evaluation device is calculated. A flicker feature amount calculation unit for calculating a flicker feature amount in each frame from the residual signal;
5. The feature quantity integration unit derives an objective evaluation value of video quality based on an average in each video sequence of the block distortion feature quantity and the flicker feature quantity. Objective video quality evaluation device.   6. The video quality objective evaluation apparatus according to claim 5, wherein the flicker feature amount calculation unit calculates the flicker feature amount based on inter-frame difference power of the residual signal.   The flicker feature amount calculation unit calculates the flicker feature amount based on an inter-frame difference power only in the residual signal where the motion vector satisfies a predetermined threshold condition. The objective evaluation device for video quality according to claim 6. A global motion vector selection unit that derives a global motion vector of each frame using the motion vector;
8. The video quality objective evaluation apparatus according to claim 1, wherein the prediction signal generation unit generates a prediction signal using the global motion vector instead of the motion vector.   9. The video quality of claim 1, further comprising a subjective evaluation value estimation unit that estimates a subjective evaluation value of video image quality by applying a predetermined nonlinear function to the objective evaluation value. Objective evaluation device.   10. The objective evaluation of video quality according to claim 9, wherein the subjective evaluation value estimation unit uses any one of a cubic function that monotonously increases or decreases and a logistic function as the predetermined nonlinear function. apparatus.