WO2005076219A1

WO2005076219A1 - Method, device, computer programme and data medium for estimating the perceptual quality of a decoded image

Info

Publication number: WO2005076219A1
Application number: PCT/FR2004/003331
Authority: WO
Inventors: Frédéric Loras
Original assignee: France Telecom
Priority date: 2004-01-09
Filing date: 2004-12-21
Publication date: 2005-08-18
Also published as: FR2865082A1

Abstract

The invention concerns a method for estimating the amount of improperly decoded data of a block Bi of an image, said image I having been coded in accordance with a predictive coding method, which consists in: estimating the amount Q1,i of unpredicted and improperly coded data of said block Bi (E1); estimating the amount Q2,i of block BI data predicted on the basis of at least one prediction data block Aj and improperly coded (E2), using the following formula: Q2,i = αji *QAj wherein QAj represents the amount of improperly coded data of the prediction block Aj and αji represents the error propagation coefficient; and in totalling the amount of improperly coded data Q1,i et Q2,i (E3) to calculate the total amount Qi of improperly coded data in the block Bi.

Description

Method, device, computer program and data carrier for estimating the perceptual quality of a decoded image

The invention relates to a method, a device and a computer program for estimating the perceptual quality of a decoded image as well as to a data medium containing the computer program. In the field of compression of still or moving images, it is known to code the images by a method of predictive spatial or temporal coding. Temporal predictive coding predicts the likely value of an element (pixel or video sequence) of an image based on its previous value and codes the difference between the two values, past and present. Spatial predictive coding predicts the likely value of an element of an image (pixel or video element) using the value of another element of this same image and codes the difference between the respective values of these two elements . Predictive coding thus allows effective compression of the image. However, it promotes the propagation of errors during decoding. Indeed, if an error is introduced in an element of the image after the coding of this one, for example during the transmission of the image between a transmitter having coded the image and a receiver able to decode it , this error spreads during decoding in all the other picture elements predicted on the basis of this erroneous element. Thus, due to the spatial and / or temporal predictive nature of the data coding method, the transmission errors spread in the image spatially and / or temporally. To hide errors, it is known to use error masking methods consisting, for example, in replacing missing or erroneous data by other correct data of the same image or of an earlier image. The masking methods make it possible to reduce the errors in the decoded data and consequently to improve the perceptual quality of the image. However, they are not completely effective. Some data remains incorrect even after being hidden. To estimate the perceptual quality of the decoded image, i.e. the quality of the image as it is perceived by a person looking at it, decoders generally measure the rate of data loss or the proportion of data not decoded in the image, without taking into account the phenomenon of propagation of the errors. The present invention aims to improve the estimation of the perceptual quality of a decoded image, the image having been coded according to a predictive coding method. To this end, the invention relates to a method for estimating the quantity of badly decoded data of a block B, of an image I, said image I having been coded according to a predictive coding method, in which: quantity Q of non-predicted and badly decoded data of said block B „the quantity Q _2ιl of data of block B„ predicted is estimated on the basis of at least one data block A, of prediction and badly decoded, using the following formula:

where Q _A) represents the quantity of badly decoded data of block A, of prediction and α ,, represents the propagation coefficient of the error, and - we sum the quantities of badly decoded data Q and Q _{2ιl in} order to calculate the quantity total Q, of badly decoded data in block B ,. The invention in fact consists in introducing a coefficient of propagation of the error in the calculation of the quantity of badly decoded data during the estimation of the perceptual quality of the image. This error propagation coefficient translates the prediction relationship between the prediction block, that is to say the block on the basis of which the prediction is made, and the predicted block. It therefore makes it possible to quantify the propagation of errors. Thanks to this, the calculation of the quantity of badly decoded data takes account of the phenomenon of propagation of the error in time and in space. Consequently, the estimate of the quality of the image thus calculated corresponds better to that actually perceived by a person viewing the image. Ultimately, the method of the invention makes it possible to take into account the fact that certain image data are not decoded correctly because they have been predicted on the basis of erroneous data. The invention also relates to a method for estimating the quantity of badly decoded data of an image I, said image I being divided into a plurality of blocks B, and having been coded according to a predictive coding method, in which, for each block B „the method according to claim 1 is executed and the quantities Q, of badly decoded data for all the data blocks B, are added. In a particular embodiment, the value of the coefficient α ,, of propagation of the error between the blocks A _j and B, is calculated using a function translating the prediction relation between the blocks A, and B ,. In another embodiment, the value of the coefficient α ,, of propagation of the error between the blocks A, and B, is between 0.2 and 0.8 when the block B, contains predicted data and data not predicted. In this case, the values of the coefficient are determined empirically. The invention also relates to a device for estimating the quantity of badly decoded data of an image I, comprising means for executing the steps of the method defined above. Preferably, the device comprises: - first estimation means arranged to estimate the quantity Q _{1ι (} of data from each block B „not predicted and badly decoded, - second estimation means arranged to estimate the quantity Q _2ιl data of each block B „predicted on the basis of at least one block of data A, prediction and badly decoded, using the following formula:

where Q _AJ represents the quantity of badly decoded data of the block A _j of prediction and ot _j , represents the coefficient of propagation of the error, - first calculation means arranged to sum the quantities of badly decoded data Q and Q ₂ , , in order to calculate the total quantity Q, of data badly decoded in block B, and - second calculation means arranged to sum the quantities Q, of badly decoded data for all the blocks B, in order to calculate the total quantity badly decoded image data. The invention also relates to a computer program comprising program instructions for executing the method previously defined by a computer as well as a data medium containing the computer program. The invention will be better understood using the following description of the method for estimating the amount of badly decoded data in an image I and of a device for estimating the amount of badly decoded data in this image I, according to a mode of particular embodiment of the invention, with reference to the drawings in which: - FIG. 1 represents a flow diagram of the steps of the method for estimating the quantity of badly decoded data of image I, according to the particular embodiment described, - FIG. 2 represents a functional block diagram of the device for estimating the quantity of badly decoded data of image I, according to the particular embodiment described and - FIG. 3 represents a transmission device, a reception device and a communication network . In the particular example of the invention, an image I is coded by a transmission device 1, according to a predictive spatial and temporal coding method, then transmitted to through a communication network 3 to a reception device 2, capable of decoding the image I. During the transmission through the network 3, certain data of the image I are lost or distorted. On reception, the device 2 here applies a known method of masking errors in order to reduce the amount of erroneous or missing data. In order to assess the perceptual quality of the image, in other words the quality of the image as it is perceived by a person looking at it, the reception device 2 performs steps 1 to 4, described below. In the particular example of the description, the image is divided into a plurality of data blocks B ,. The decoding of each data block B uses two types of data: - non-predicted data, coming directly from the network 3, and - predicted data, reconstructed from one or more prediction data blocks belonging to the same image (in this case it is a spatial prediction) or to an earlier image (it is a time prediction in this case). To estimate the quantity of badly decoded data in each block of data B, the device 2 first performs the following steps 1 to 3: Step E1: the device 2 estimates the quantity Q ^ of data in block B, not predicted and incorrectly decoded using the following formula: Ql, = β * Q. max = (Pl * β2) * Q. max where Q, _max represents the maximum quantity of badly decoded data of block B „here equal to the total quantity of data of block B„ p _! represents the proportion of data in block B, decoded without prediction, on the basis of data coming directly from the network 3, and β ₂ represents the proportion of badly decoded data among these data decoded without prediction (namely among the β ₂ ^* Q, _max data). If all the data used to decode the block B, come from the network, the coefficient β-i is equal to 1. If all the data coming from the network are correct, the coefficient β ₂ is equal to 0. On the other hand, if all the data from the network is lost, β ₂ is equal to 1. If only half of the data from the network is correct, β ₂ is equal to Yz. Step E2: the device 2 estimates the quantity Q _2ιl of data of block B „predicted on the basis of at least one block of data A, of prediction and badly decoded, using the following formula:

Q _AJ is the quantity of badly decoded data of block A, of prediction and α, is the propagation coefficient of the error. The quantity Q _{2 | l} represents the wrong data portion decoded from block B, due to the phenomenon of propagation of the error during decoding. The coefficient α ,, of propagation of the error depends on the prediction relation between the prediction block A, and the predicted block B ,. In the case where the block B, is totally predicted on the basis of the prediction block A ,, α ,, is worth 1, the dependence between A, and B, being total In the opposite case where the block B, is not predicted on the basis of block A ,, α ,, is worth 0, no prediction dependence existing between blocks A, and B ,. In the case where block B, is partially predicted on the basis of block A ,, the coefficient α ,, is between 0.2 and 0.8, its value being determined empirically. An empirical method for evaluating the value of the coefficient α ,, consists in making the posteriori link between the calculated values Q _Aj of quantities of badly decoded data of the prediction blocks A, and the overall perceptual quality of the reconstructed image. It is possible, for example, to subjectively assign ratings to the quality of the reconstructed images and retrospectively determine the value of α ,, which would make it possible to obtain the values of Q _Aj closest to the previously assigned ratings. A value of coefficient α ,, less than 1 translates the dispersion of the error in time or in space thanks to the correct data to which the erroneous data coming from block A, are added. In addition, in the case where, as in the present case, the decoding device 2 masks the errors, the value of the coefficient α ,, partially depends on the effectiveness of the error masking method used. The more efficient this method, the lower the value of the coefficient α. In the particular example of the description, an error masking method being applied, the value of the propagation coefficient α, is fixed at 0.7. Experiments have shown that this value gives good results. In the case where block B, is predicted on the basis of several prediction blocks A _J? the quantity Q _2ιl of badly decoded data, due to the propagation of the error, is equal to: - Aj

Step E3: the device 2 sums the quantities of badly decoded data Q _1fl and Q _{2ιl in} order to calculate the total quantity Q, of badly decoded data in block B,. During a step E4, the device 2 sums the quantities Q, of badly decoded data for all the data blocks B, in order to calculate the total amount of badly decoded data in the image I. The reception device 2 , shown in FIG. 2, comprises a decoding module 20, a module 21 for masking errors and a module 22 for estimating the perceptual quality of an image. The decoding module 20, supplied with input by the coded image data received by the network 3, is intended to decode this image data. It is recalled here that, due to the transmission, certain image data, received by the device 2, may be missing or erroneous. The error masking module 21, connected as an input to an output of the decoding module 20, is arranged to mask this erroneous data, at least partially, using a known error masking technique. The module 22 for estimating the perceptual quality of an image comprises a first estimation block 220, arranged to estimate the quantity Q of data of each block B „not predicted and poorly decoded, and a second estimation block 221 , arranged to estimate the quantity Q _2pl of data of each block B „predicted on the basis of at least one block of data A, prediction and badly decoded. The amounts of badly decoded data Q _1§ , and Q _2τl are calculated using the formulas previously indicated in the description of the process. The estimation module 22 also includes: a first calculation block 222, connected to the two estimation blocks 220 and 221, arranged to sum the quantities of badly decoded data Q and Q _{2] l in} order to calculate the total quantity Q , badly decoded data in block B, and - a second calculation block 223, connected to the first calculation block 222, arranged to sum the quantities Q, of badly decoded data for all the blocks B, in order to calculate the total quantity Q of badly decoded data of image I. The invention which has just been described relates to a method and a device for estimating the quantity Q of badly decoded data of image I. In a preferred embodiment, the process steps are executed by a computer under the command of program instructions. Consequently, the invention also relates to a computer program intended to be stored in or transmitted by a data medium comprising program instructions for executing the method by a computer as well as the data medium on which the program is implemented . The data medium can be a hardware storage medium, for example a CD-ROM, a magnetic diskette or a hard disk, or else a transmissible medium such as an electrical, optical or radio signal, this signal being able to be transmitted through an electrical, optical or aerial cable. In the above description, the value of the coefficient α ,, of propagation of the error between the blocks A _j and B, is determined empirically, taking into account the respective proportions of the predicted data and of the non-predicted data coming directly from the network and the effectiveness of the masking method. In another embodiment, the value of the propagation coefficient α ,, is calculated, using a function translating the prediction relation between blocks A, and B ,. As a nonlimiting illustrative example, we will now give a particular example of a function for calculating the coefficient a _β translating the prediction relationship between the blocks Aj and B, within the framework of the H.261 video compression standards (ITU Recommendation -T H.261). This function for calculating the propagation coefficient α ,, is also applicable to the video compression standards H.263 (ITU-T Recommendation H.263 rev2) and MPEG-4 Part 2 (ISO / IEC / 14496/2 MPEG-4 visual ). The propagation coefficient a _β is calculated as follows: - if there is no dependence between the prediction block A _j and the predicted block B „then _β = 0 otherwise rt = 1 N ^'• b ~ Bitrans

where Nb _Bllrans represents the number of data in block B, coming directly from the network, in other words the non-predicted data in block B „and Nb _{Λ max} represents the maximum number of data that block B can contain. Note in the video compression standards “ITU-T H.261”, “ITU-T H.263” and “ISO / IEC / 14496/2 MPEG-4 visual”, the quantity of data Nb _Bllrans is represented by the number of coefficients "TCOEFF" of the current block and Nb _{Bι max} is represented by the maximum number of "TCOEFF" that it is possible to transmit for this block. As an illustrative example, we will explain two particular cases of calculation of the propagation coefficient a _β . In the first case, no data specific to block B, comes from the network. It follows that Nb _Bttrms = 0 and therefore α, = 1 - The predicted block B, is totally dependent on the prediction block A _j . In a second case, all the data in prediction block B, come from the

network: = 1 ^{? JIU≡} - = 1 - = - • The data of B, are added to the data of 2 χ Nb _{fl (max} 2 2

_{Up to date} The proportion of dependence of B, on A, is then ¹ A It will be emphasized that the calculation of the propagation coefficient a _β depends intimately on the way in which this prediction is made. For the H.264 standard (ITU-T H.264), it would have to be calculated in a different way. In this case, the value of the propagation coefficient is determined more precisely but requires much larger calculations.

Claims

claims

1. Method for estimating the quantity of badly decoded data of a block B, of an image I, said image I having been coded according to a predictive coding method, in which: - the quantity Q of non-predicted data is estimated and badly decoded from said block B, (E1), - the quantity Q ₂ ,, of data of block B „predicted on the basis of at least one data block A _j of prediction and badly decoded (E2) is estimated at l using the following formula:

where Q _AJ represents the quantity of badly decoded data of block A, of prediction and cc _j , represents the coefficient of propagation of the error, and - we sum the quantities of badly decoded data Q and Q ₂ ,, (E3) so to calculate the total quantity Q, of data badly decoded in block B ,.

2. Method for estimating the quantity of badly decoded data of an image I, said image I being divided into a plurality of blocks B, and having been coded according to a predictive coding method, in which, for each block B „, we execute the method according to claim 1 and the quantities Q, of badly decoded data for all the data blocks B, (E4) are added.

3. Method according to one of claims 1 and 2, wherein the value of the coefficient α ,, of propagation of the error between the blocks A _j and B, is calculated using a function translating the relation of prediction between blocks A, and B ,.

4. Method according to one of claims 1 and 2, wherein the value of the coefficient α ,, of propagation of the error between the blocks A, and B, is between 0.2 and 0.8 when the block B , contains predicted data and non-predicted data.

5. Method according to one of claims 1 to 4, in which, during step (E2) of estimating the quantity Q ₂ ,, of data of block B „in the case where the data of block B, are coded on the basis of a plurality of prediction blocks A ,, the quantity Q _2ιl of data is equal to:

6. Device for estimating the quantity of badly decoded data of an image I, comprising means for executing the steps of the method according to one of claims 2 to 5.

7. Device according to claim 6, in which it is provided:

- first means of estimation (220) arranged to estimate the quantity _Qi,, data of each block B "not predicted and decoded bad,

- second estimation means (221) arranged to estimate the quantity Q _{2 | l} of data of each block B „predicted on the basis of at least one data block A _j of prediction and badly decoded, using the following formula: Q _2l , = a _j * QA, where Q _Aj represents the quantity of badly decoded data of block A, of prediction and α ,, represents the propagation coefficient of the error,

first calculation means (222) arranged to sum the quantities of badly decoded data Q ^ and Q _{2ιl in} order to calculate the total quantity Q, of badly decoded data in block B, and

- second calculation means (223) arranged to sum the quantities Q, of badly decoded data for all the blocks B, in order to calculate the total quantity of badly decoded data of the image.

8. Computer program comprising program instructions for executing the method according to one of claims 2 to 5 by a computer.