WO2010007286A2

WO2010007286A2 - Method of referenceless measurement of the perceived quality of a signal and corresponding device

Info

Publication number: WO2010007286A2
Application number: PCT/FR2009/051186
Authority: WO
Inventors: Nicolas Ramin; Ricardo Pastrana Vidal
Original assignee: France Telecom
Priority date: 2008-06-24
Filing date: 2009-06-22
Publication date: 2010-01-21
Also published as: WO2010007286A3

Abstract

The invention relates to a method of referenceless measurement of the perceived quality of a signal (S) carrying data representative of an at least visual or audible content, characterized in that it comprises the steps of: indexation (E2) of said content, determining at least one class (C) of semantic membership of said content, and estimation (E4) of the perceived quality of said signal (S) by using a method of referenceless measurement of perceived quality, said method being adapted to said membership class (C) or to a type (t) of content of said class (C).

Description

Measurement method without reference to the perceived quality of a signal and corresponding device

The present invention relates generally to the field of signal processing, and in particular to the evaluation of the subjective quality of image, video, audio or audiovisual signals.

The invention relates to a method of objective evaluation of the subjective quality, that is to say as perceived by a recipient, such signals. Given an initial non-degraded signal, it is sought to evaluate the subjective quality of this signal when it is degraded due to transformations or transmission through a communication chain. The invention thus finds particular application in the monitoring, control, adaptation or optimization of communication channels, at the coding, transmission or playback of the degraded initial signal.

It is recalled here that objective measurements of the quality of visual or sound signals are automated measurements that oppose subjective measurements, made visually or audibly by natural persons. Subjective measures of quality remain the most used and most effective measures in the field of measuring the quality of audio or image signals, which include video signals, but are very expensive since they require to be put in place. a panel of people in situations of listening or seeing sound or visual samples of the signals to be evaluated.

This is why the development of alternative methods of objective measurement that can supplement or supplement subjective methods is a subject of great interest. The existing methods of objective measurement of the quality of an audio or image signal are mainly divided into three categories: The first category brings together fully referenced measurement methods, which compare the degraded signal to be evaluated with the corresponding initial signal. One of these methods is for example the measurement of the PSNR (according to the English "Peak Signal-to-Noise Ratio") or even the method BTFR (according to English "British Telecom FuII

Reference ") described in Annex A of the standardization document [ITU-T J.144] of I ¹ ITU (from the English" International Telecommunication Union ") entitled" Objective perceptual video quality measurement technology for digital cable TV in the presence of a full reference "in 2004. The use of such methods has the disadvantage of requiring the availability of the initial signal, which is not always possible especially when measuring the quality of the degraded signal at the end of The second category concerns the measurement methods with reduced reference, which use information extracted from the initial signal and the degraded signal to be evaluated, to measure the subjective quality of this signal. These methods compare, for example, statistical data calculated on the initial signal, such as a measure of blur or a measure of block effects on the initial signal by e. xemple, with corresponding statistical data calculated on the degraded signal to be evaluated. Such a method is described in the IEEE article entitled "A Reduced-Reference Perceptual Quality Metric for In-Service Image Quality Assessment" by T. Maulana Kusuma et al. ., published in 2003. Such a method has also been proposed by the NTIA (National Telecommunications and Information Administration) and is described in Annex D of the ITU-T J.144 ] pre-cited. This second category of objective measurement methods requires the availability or calculation of information on the initial signal, which is not applicable in some contexts of quality measures, particularly at the end of the communication chain, when transmitting live audiovisual content, for example.

- Finally, the third category brings together measurement methods without reference, which evaluate the quality of a degraded signal without using either the corresponding initial signal or information extracted from this initial signal. These methods focus on measuring events on the degraded signal to be evaluated, such as the amplitude and duration of signal degradations, to evaluate the perceived quality of the degraded signal. Such a method is described in the patent application FR2884087 to R. Pastrana et al., Entitled "Method and device for assessing quality degradation caused by invariance of a stimulus, as perceived by a recipient of said stimulus". These non-reference measurement methods thus offer the advantage of being usable on all the elements of a communication chain in which the degraded signal to be evaluated passes, without requiring an additional channel conveying information on the corresponding initial signal.

However, current methods of objective measurement without reference do not predict with robustness and accuracy the opinion of users as to the perceived quality of the audio signal or image to be evaluated. In particular, one of the difficulties of these methods consists in distinguishing the true defects of the degraded signal from the content of the initial signal, which humans are able to perform from context and experience.

One of the aims of the present invention is to overcome the disadvantages of the prior art by providing a method and a measuring device without reference to the perceived quality of a signal carrying data representative of a content at least visual or sound. To this end, the invention proposes a measurement method without reference to the perceived quality of a signal carrying data representative of a at least visual or sound content, characterized in that it comprises the steps of:

indexing said content, determining at least one semantic class of said content, and estimating the perceived quality of said signal by using a measurement method without reference of perceived quality, said method being adapted to said class of membership or a content type of said class.

In fact, the inventors have noted that the semantics attached to the different classes or types of content have an influence that is not taken into account by the current measurement methods based on the only signal analysis. For example, a signal degradation does not have the same impact on a video showing an athletic race as on a more static type of video such as an interview. By taking into account the semantic content of the signal whose quality is evaluated, the invention thus makes it possible to maximize the accuracy and robustness of the measurement without reference to the perceived quality of this signal, depending on the application context. The device and the method according to the invention strongly restrict the current inter-content dispersion of the estimates without reference to perceived quality.

According to an advantageous characteristic, said estimation step is preceded by a step of selecting said measurement method without reference to perceived quality among a set of non-reference measurement methods adapted to various types or classes of content semantic membership.

This feature makes it easy to integrate new measurement methods without reference adapted to certain types of content, in a measurement device without reference to the perceived quality according to the invention. According to another advantageous characteristic, said selection step takes account of at least one speed parameter of the measurement method to be selected or a parameter of measurement accuracy of perceived quality.

This feature makes it possible to use the non-reference measurement method according to the invention in very different contexts: for example to evaluate the perceived quality of a signal in a piece of equipment of a communication chain, an expensive measurement method is selected. in calculations of which quite slow but very precise, while to evaluate the perceived quality of a signal in a mobile terminal, of low capacity, one selects a method of measurement inexpensive in computing resources and therefore fast even if less precise. The method according to the invention is thus portable depending on the context, but also available data. Thus, if meta-information, not extracted from the initial source signal but bearing for example on the nature of the signal to be evaluated, or on a specific application context, are available at the level of the measuring device without reference according to the invention, these meta-information information is used to further improve the accuracy of the measurement without reference to the perceived quality of the signal.

According to another advantageous characteristic, said indexing step is preceded by a step of selecting said indexing method from among a set of indexing methods, said selection step taking account of at least one speed parameter of the indexing method. indexing method to be selected or a measurement accuracy parameter of perceived quality.

The choice of the indexing method used according to the context and the available data further contributes to the portability of the measurement method without reference to the perceived quality of a signal according to the invention.

According to another advantageous characteristic, said indexing step takes account of at least one meta-information on said signal.

This characteristic makes it possible to directly index the signal to be evaluated by using a meta-information on this signal, for example an indication of the semantic content of the signal, without using an indexing method. complex using a motion analysis or colors if it is a video signal for example.

According to another advantageous characteristic, said estimation step comprises a substep of extraction in said signal of primitives related to the estimation of the perceived quality of said signal, said primitives being adapted to said membership class or said type. content of said class.

This sub-step of extraction of primitives, common to most existing measurements without reference of perceived quality, is here adapted to the class or the type of content of the signal to be evaluated: one extracts only the primitives of the signal which are relevant for evaluate the perceived quality of this signal given the class or type of the corresponding content. This adapted extraction of primitives makes it possible to further improve the accuracy and the robustness of the estimate of the perceived quality of the signal. It also makes it possible to save computing resources with respect to a quality measurement that generically extracts any primitive that is useful for measuring the quality of any content, and then takes into account the type of content corresponding to the signal to be evaluated only when a last step of integration of the signal degradation measures to arrive at a score of perceived quality.

According to another advantageous characteristic, when said signal carries representative data of one or more images, said estimation step comprises a substep of analysis of at least one of said images, and a substep of extraction. primitives related to the estimation of the perceived quality of said signal as a function of one or more regions identified in said image during the analysis step. When the signal to be analyzed is a video, it is indeed useful to analyze the various scenes that compose it in order to adapt the measurement of the perceived quality of the signal to these different scenes, each of these scenes being of a type. different content. For example, for video content showing a tennis match, there are images showing the tennis court of those showing a close-up player to measure the quality differently on these different images. On the images showing the field of tennis, we identify different regions including the terrain and the public. Then we extract sharpness measurements only on the part of the terrain in these images, and not on the part corresponding to the public, on which a recipient of the signal does not expect sharpness. This prior analysis of the images carried by the signal to be evaluated therefore makes it possible to predict in a very precise manner the quality perceived by a recipient of the signal, as a function of very specific contents.

The invention also relates to a measurement device without reference to the perceived quality of a signal carrying data representative of a content at least visual or sound, characterized in that it comprises:

means for obtaining data for indexing said content, determining at least one semantic membership class of said content,

and means for estimating the perceived quality of said signal by using a measurement method without reference of perceived quality, said method being adapted to said class of membership or to a type of content of said class.

It should be noted that the means for obtaining indexing data are limited in certain cases to sending a request to an indexing server and receiving a response from this indexing server, for example. example when the indexing of the content is performed outside the device according to the invention in this indexing server.

The invention also relates to a terminal incorporating the device according to the invention, as well as a computer program comprising instructions for implementing the measurement method without reference according to the invention, when it is executed on a computer.

The non-reference measuring device according to the invention, as well as the computer program, have advantages similar to those of the method according to the invention. Other features and advantages will appear on reading a preferred embodiment described with reference to the figures in which:

FIG. 1 represents a measurement device without reference to the perceived quality of a signal according to the invention, used at the end of a communication chain delivering the signal,

FIG. 2 represents steps of the measurement method without reference to the perceived quality of a signal according to the invention,

FIG. 3 represents an embodiment of the measuring device without reference of perceived quality according to the invention; FIG. 4 represents sub-steps of a step of estimation of perceived quality of the method according to the invention, when said estimate has low computing resources,

and FIG. 5 represents sub-steps of a perceived quality estimation step according to the invention, when said estimate has larger calculation resources.

According to a preferred embodiment of the invention, it is implemented in a device G shown in Figure 1, placed at the end of a communication chain. The latter comprises at its source a device DP for producing a content, which is delivered to a coding device DC which encodes this content in an initial signal transmitted over a communication network RES to a decoding device DD. The decoding device DD delivers a signal S that is degraded with respect to the initial signal, to the measuring device G without reference to perceived quality according to the invention. It should be noted that in this embodiment of the invention, the signal S corresponds to an image, audio, video or audiovisual content, the principle of the invention remaining the same in each of these cases, and certain process steps. according to the invention being detailed case by case when this is necessary for a good illustration of the invention. The device G is for example integrated in a terminal rendering the content corresponding to the degraded signal to a user. It is of course usable at other locations of the communication chain, for example in a decoder placed directly at the output of the DC encoding device.

The device G receives as input the degraded signal S, a measurement parameter p of perceived quality and a parameter k of measurement method or indexing speed used by the method. The parameters p and k make it possible to adapt the measurement method without a reference of perceived quality according to the invention, implemented in the device G, to a given application context: for example if the device G is implemented in a server controlling the quality of the device. the communication chain, the parameter p will indicate a desired maximum quality measurement accuracy and the parameter k a possible large calculation time of measurement or indexing method; on the other hand, if the device G is implemented in a mobile terminal, and therefore has low computing and display capacities, the parameter p will indicate a measurement accuracy of average quality and the parameter k a rapid calculation speed of measurement method or indexing.

The device G also optionally receives as input a parameter m comprising one or more meta-information on the signal S, not extracted from the initial signal, depending on the availability of these meta-information. For example, these meta-information items are an indication of the user service related to the signal S, an indication of the nature of the content corresponding to the signal S, or an indication of the quality of the transmission of the signal S. These indications are in fact susceptible of accelerate or modify certain steps of the measurement method without reference to perceived quality according to the invention, as detailed below.

At the output, the device G delivers an estimate Q of the perceived quality of the signal S, on a measurement scale of perceived quality. In this embodiment, this estimate Q is delivered on a scale of 0 to 100, the minimum of this scale indicating a very poor perceived quality and the maximum of this scale indicating a very good perceived quality. In variant is used an MOS score (according to the English "Mean Opinion Score") of the signal S, varying on a scale of 1 to 5, or the ACR scale (of the English "Absolute Category Rating") defined in the "ITU-T Recommendation P. 910" entitled "Subjective Video Quality Assessment Methods for Multimedia Applications", or the SAMVIQ ("Subjective Assessment Methodology for Video Quality"), defined in Recommendation ITU-R BT.1788.

It should be noted that in this embodiment of the invention, the scale used is the same irrespective of the value of the precision parameter p supplied at the input of the device G. Indeed, only the reliability of the estimation of the Quality measurement varies according to the precision p parameter.

With reference to FIG. 2, the perceived quality of the signal S is estimated in four steps E1 to E4 by the measurement method without reference to perceived quality according to the invention implemented in the device G.

Step E1 is the selection of an indexing method from among a set of indexing methods, as a function of the precision parameter p and the speed parameter k. This step E1 is implemented in a software module G1 of the device G, represented in FIG.

In this step E1, the software module G1 selects either an indexing method by type of content, implemented in a software module G21, or a content class indexing method, implemented in a software module G22, the method of Indexing by content class is a coarser indexing method and therefore less complex than the indexing method by type of content. Indeed a content class is a content semantic membership class, encompassing several types of content. For example, for a signal S corresponding to a video content, there is for example, among other classes of content, a class encompassing the video content of grass sports, a class encompassing the video content of track sports, a class encompassing sky-sea documentaries, and a class encompassing videos showing a presenter or interlocutor. In the class encompassing grass sports videos, there are several types of content, such as a content type encompassing football match videos, a content type encompassing rugby videos, or a content type encompassing a football game. polo match.

Thus, in this step E1, if the parameter p and the parameter k indicate a desired maximum accuracy of measurement of perceived quality, or a possible significant calculation time, the software module G1 selects the indexing method by type of content, and sends the signal S to the software module G21. On the other hand, if the parameter p indicates a low accuracy of perceived quality measurement and the parameter k a fast calculation speed, the software module G1 selects the indexing method by content class, and sends the signal S to the software module G22 .

The next step E2 is the indexing of the content corresponding to the signal S. This step is implemented in the software module G2, and more precisely either by the software module G21, or by the software module G22, depending on the method of indexing selected in step E1. When a method of indexing by content class has been selected in step E1, and a parameter m containing information on the nature of the content corresponding to the signal S is available at the input of the device G, then in this step E2 the indexing of the content corresponding to the signal S uses this parameter m. If the signal S is a video, this parameter m is for example obtained by interrogating a video broadcast programming server, and indicates a content class or a content type, which do not necessarily correspond to the classification of the contents. used by the device G for their indexing. In the latter case the module G22 uses a correspondence table between the classification of the programming server and that used by the device G, to determine a content class C corresponding to the signal S. On the other hand, when a method of indexing by content class has been selected in step E1, and a parameter m containing information on the nature of the content corresponding to the signal S is not available at the input of the device G, in this step E2 the module G22 uses a coarse indexing method for classifying the content of the signal S in a content class C.

For example, if the signal S is a video, a colorimetric study of the images of the signal S, using for example a distance between a dominant color in these images and the dominant colors associated with the content classes defined in the device G, allows, in association with a motion estimation of the camera having taken these images, to roughly index the video corresponding to the signal S. A method for estimating camera movement that can be used is described in the article "Robust Multiresolution Estimation of Parametric Motion Models" of J. Odobez et al., Published in 1995 in the Journal of Visual Communication and Image Representation. So:

- if the camera has a low amplitude motion and the images have a dominant color similar to that of the flesh of a face, the video corresponding to the signal S is classified in the content class including the videos showing a presenter or interlocutor, if the camera has a large amplitude movement and the images have a dominant gray or brown color, the video corresponding to the signal S is classified in the content class including the track sports videos,

and if the camera has a large amplitude movement and the images have a dominant green color, the video corresponding to the signal S is classified in the content class encompassing the grass sports videos. If the signal S is an audio signal, the module G22 performs, for example, a rough indexing of the signal S in a content class corresponding to the speech signals, or in a content class corresponding to the music signals. This indexing uses, for example, models of sound signals corresponding to music or speech, to classify the signal S. Indexing methods that can be used for the sound signals in this module G22 are described in:

the article by J. A. Marks et al. entitled "Real time speech classification and pitch detection" and published in 1988 on the occasion of an international conference "Communications and Signal Processing" that took place the same year, - or the article by Y. Zhu and D. Zhou entitled "Scene changes detection based on audio and video content analysis", and published in 2003 for the fifth international conference "International Conference on Computational Intelligence and Multimedia Applications". Similarly, when a content type indexing method has been selected in step E1, and a parameter m containing information on the nature of the content corresponding to the signal S is available at the input of the device G, then in this step E2 the module G21 uses this parameter m to determine a type t of content corresponding to the signal S. If on the other hand such a parameter m is unavailable or if the information contained in this parameter m is insufficient to determine a type of content associated with the signal S, then the module G21 uses in this step E2 a fine indexing method for classifying the content of the signal S in a type t of content, within the limits imposed by the parameter k indicating the maximum possible calculation time for the measurement of perceived quality of the signal S.

For example, if the signal S is a video, the module G21 uses the indexing method of the module G22 to determine a class C of content associated with the signal S, then uses forms recognition methods specific to the content types of the class C for indexing the content corresponding to the signal S in a type t of content. For example if class C is the class encompassing grass sports videos, module G21 detects the shape of the goals or the layout of the field lines in the images of the signal S to determine if it corresponds to the type of content encompassing the match videos football or content type that includes rugby match videos.

Similarly, if the signal S is an audio signal, the module G21 uses the indexing method of the module G22 to determine a class C of content associated with the signal S, then uses, for example, observations functions that are specific to certain types of content. audio to determine the type t of content corresponding to the signal S. Such observation functions are described for the class of musical signals, in the thesis of S. Rossignol of the University of Paris 6, entitled "Segmentation and indexation of musical sound signals ", and supported in 2000.

When the module G21 fails to determine a type t of content corresponding to the signal S, for example because such observation functions are not available for the class C content associated with the signal S, or because the The maximum computation time imposed by the parameter k does not allow a fine indexing of this content, the module G21 merely indexes the content of the signal S per content class. At the end of step E2, the module G21 indexed the content corresponding to the signal S in a content type t or a content class C, or the module G22 indexed the content corresponding to the signal S in a class C of content. The type t of content or the class C of content are output from the modules G21 or G22 to the software module G3 for example in the form of integers, each assigned to a content type or a specific content class in the classification. contents used by the device G.

Alternatively, the content type t or the content class C are provided to the software module G3 in the form of probability vectors. For example, the probability vector associated with the type t of content indicates the probability of belonging of the content corresponding to the signal S to each of the types of content. contents belonging to the classification of contents used by the device G.

The next step E3 is the selection of a measurement without reference of perceived quality, adapted to the content class C or the content type t corresponding to the signal S and determined at the end of the step E2. This step is implemented in the G3 software module, and takes into account the precision p and k speed parameters.

In this step E3, the software module G3 sends the signal S, ie: to the software module G41, which adapts a measurement method without reference of perceived quality to the type t or to the content class C determined in step E2, by a specific extraction of primitives and / or a specific parameterization of this method, either to the module G42, which uses a measurement without reference of perceived quality more complex and more precise, also adapted to the type t or the class C of contents determined with step E2, but using a preliminary analysis of the signal S. Thus for example in this step E3:

if the parameter p and the parameter k indicate a desired maximum accuracy of measurement of perceived quality, and a possible large calculation time, the software module G3 selects the measurement without reference of perceived quality implemented by the software module G42, and sends the signal S to the software module G42 if no measurement implemented in the software module G41 makes it possible to accurately predict the quality of the signals of the same semantic content, if the parameter k indicates a desired low calculation time, the software module G3 selects the measurement without perceived quality reference implemented by the software module G41, and sends the signal S to the software module G41, - and if the parameter p indicates a low measurement accuracy of perceived quality, and the parameter k a significant calculation time possible, the software module G3 selects the measurement without a perceived quality reference implemented by the software module G41, and sends the signal S to the software module G41.

The next step E4 is the estimation of the perceived quality of the signal S, without reference to the corresponding initial signal. This step is implemented in the software module G4, and more precisely either in the software module G41 or in the software module G42, according to the method without reference selected in step E3.

If in step E3 the software module G3 has sent the signal S to the module G41, then the estimation of the perceived quality of the signal S takes place in three steps E41A to E43A shown in FIG. 4:

Step E41A is a primitive extraction step related to the estimation of the perceived quality of the signal S.

As a function of the signal S and of the type t or of the class C of the content associated with the signal S, primitives of the signal S are extracted in this step E41A. Primitives that can be used to measure the perceived quality of the signal S are, for example: image or image sequence of the spatial gradients of the signal S if the signal S is an image or a video,

the average movement of the pixels between two consecutive images of the signal S, the duration and the number of image gels of the signal S or of block effects on the signal S if the signal S is a video,

and the duration and the number of signal losses if the signal S is an audio signal.

As a function of the method without a perceived quality measurement reference used, this primitive extraction step E41A is generic or, on the contrary, adapted to the class C or the type t of the content corresponding to the signal S. This adaptation consists, for example, in a choice of primitives to extract or in a specific way to extract them. For example, in the case where the signal S is an audio signal, this step E41A extracts the number of signal losses whose duration is greater than a certain threshold, this threshold depending on the class C or the type t of the corresponding content. to the signal S. Indeed the perceptual detection threshold of a signal loss is slightly different if the signal considered is a speech signal or a musical signal. Similarly, the perceptual detection threshold of a freeze of images is different depending on whether the corresponding video shows a documentary, a newscast or sports content. Step E42A is the calculation of the degradations on the signal S, using the primitives extracted in step E41A. For example, if the signal S is an audio or video signal, and if in step E41 A:

detects the signal losses in the signal S, which result in an audio cutoff and / or an image freeze for a video signal,

calculated their durations and filtered the losses detected to select only those whose durations are greater than a threshold of perceptual detection,

and then calculating the temporal densities of the signal losses as a function of their durations, then in this step E42A, using the time densities of the signal losses thus calculated, a cognitive model is applied to these signal losses, giving an estimate of the degradation of the subjective quality of the signal due to loss of signal. This cognitive model is different depending on whether one calculates a degradation due to the losses of audio signal, or a degradation due to the freezes of images. It uses a temporal accumulation of impairments due to signal losses in the audio or video signal sequence. Moreover, its parameters are determined in such a way as to obtain a strong correlation between measurement results of perceived quality derived from subjective tests and measurement results of perceived quality provided by the cognitive model. Such a cognitive model is described for video signals in the article by Ricardo R. Pastrana et al., Entitled "A no-reference video quality metric based on a human assessment model", and published in 2007 on the occasion of the third International Workshop "Video Processing and Quality Metrics for Consumer Electronics (VPQM)". This model is adaptable to audio signal losses.

Step E43A is the integration of the degradation calculations in step E42A into a cognitive model giving an estimate of the perceived quality of the signal S, taking into account all the types of impairments taken into account in step E42A. This integration is adapted to the class C or the type t of the content corresponding to the signal S, by integrating degradations themselves calculated in a manner adapted to the class C or the type t of this content and / or by the use a cognitive model parameterized in a way adapted to the class C or the type t of this content.

For example, for a video signal S, the cognitive model described in the patent application EP1864510 is used, which makes it possible to give an estimate of the perceived quality of a video signal taking into account the interaction between the spatial degradations and the degradations. The temporal impairments correspond, for example, to the image gels detected on the image sequence corresponding to the signal S, by using the cognitive model described above in relation with the step E42A. The spatial degradations correspond for example to measurements of blur, carried out with the method without reference described in the article of P. Marziliano and al., Entitled "A no-reference perceptual blur metric", and published in 2002 on the occasion an international conference "International Conference on Image Processing". In this step E43A, a cognitive model is used whose parameters are determined in such a way as to obtain a strong correlation between measurement results of perceived quality derived from subjective tests on contents of class C or of type t determined in step E2. , and perceived quality measurement results provided by the cognitive model. The module G41 has for this cognitive model, a set of parameters by content type and by content class belonging to the content classification used by device G.

Other cognitive models are of course usable in this step E43A, for example a model integrating the contribution of other types of impairments, for example by means of block effect measurements carried out at step E42A when the signal S is a video.

If in step E3 the software module G3 sent the signal S to the module G42, then in step E4 the estimation of the perceived quality of the signal S uses a preliminary analysis of the signal. This analysis makes it possible to extract low or high level primitives associated with certain components of the signal, for which the impact of the impairments on the perceived quality of the signal S is different according to the nature of these components. For example, for a signal S corresponding to a video signal, the estimation of the perceived quality of the signal S takes place in five steps E41 B to E45B represented in FIG.

Step E41 B is the analysis of an image of a sequence of images of the signal S. This analysis makes it possible to detect regions of interest in the signal S, as a function of the class C or the type t of content determined in step E2. For example, if in step E2 it has been determined that the content of the signal S belongs to the class of content encompassing the videos showing a presenter or interlocutor, one or more faces are detected in this image. This detection uses, for example, the localization method described in the article by C. Garcia and M. Delakis, entitled "Convolutional Face Finder: a Neural Architecture for Fast and Robust Face Detection" and published in the IEEE magazine (based on Institute of Electrical and Electronic Engineer "Transactions on Pattern Analysis and Machine Intelligence" in November 2004.

Similarly, if in step E2 it has been determined that the content of the signal S belongs to the content class encompassing the sky-sea documentaries, in step E41 B, one or more regions corresponding to FIG. sky and one or more regions corresponding to the sea. For this, the regions of the image having a colorimetry close to that of the sky or the clouds are detected, and the regions of the image having a colorimetry close to that of the sea, by using a set of reference images for these colorimetric determinations. The impact of the degradations on the regions thus detected will not be the same, from the point of view of the perceived quality of the signal S, as the impact of the degradations on the other regions of the image, which correspond to objects attracting the attention of a viewer.

It should also be noted that in this step E41 B, several successive or regular images are preferably analyzed in the signal S in order to detect changes in the content of the signal S. For example, this makes it possible to detect that a sequence of images showing a presenter is followed by a sequence of images with the presenter and an interlocutor.

The following step E42B is the adapted extraction of primitives useful for the calculation of the impairments on the signal S, impacting the perceived quality of the signal S. These primitives are for example the loss of an object of interest in one or more images signal S, or the detection of block effects or blur on this object of interest.

In the case where the signal S is a sky-sea documentary video in the images of which one or more sky regions and one or more sea regions have been detected in step E41 B, for example, this step E42B is extracted. :

the density of false boundaries, as well as the density of homogeneous zones, in the sky regions, the density of false boundaries, the density of homogeneous zones, and a quantification of image effects due to discrete cosine transforms , in the sea regions,

and the density of false boundaries, as well as the density of homogeneous zones, on the other regions of the image. This extraction of primitives uses for example the works of Z. Wang et al., Published in 2002 in the article "No reference perceptual quality assessment of jpeg compressed! images "on the occasion of an international conference" International Conference on Image Processing ".

Step E43B, performed after or in parallel with this step E42B, is the selection of an ad hoc cognitive model as a function of the analysis carried out in step E41B and according to the class C or the type t of content corresponding to the signal S. This cognitive model is more particularly adapted to the regions of interest detected in step E41 B: for example, in the case where the signal S is a sky-sea documentary video, this cognitive model takes into account the degradations due to the compression effects of the images, differently depending on the regions of degraded images. This cognitive model is for example a model specific to a type t or a class C of content, in which the impacts of each type of degradation are weighted according to the regions impacted by these impairments in the images of the signal S. For a class C content or a type t of content and a combination of regions of interest detected in step E41 B, is determined in this step E43B an ad hoc cognitive model. This model gathers at the same time:

the information necessary for calculating the degradations of a semantic object according to the primitives extracted in step E42B,

and the information necessary for estimating the overall perceived quality of the signal S from the degradations thus calculated, this information being previously determined by subjective tests.

Thus, in the case where the signal S is a sky-sea documentary video, this cognitive model is for example:

Q is the estimate of the perceived quality of the signal S, is the sum of the degradations taken into account in this

model, clipped to a perceptual maximum,

- d, is the sum of the degradations considered on one of the semantic objects detected in the signal S, that is to say either on all the sky regions, or on all the sea regions, or on the other regions of the S signal images,

S ₁ is a weighting coefficient of the impairments considered on this semantic object, this coefficient being previously determined by subjective tests, P _1J ^{is a} primitive extracted in step E42B on this semantic object, for example the density of zones. homogeneous in the sky regions if the semantic object considered is the sky,

and _y is a weighting coefficient of this primitive, previously determined by subjective tests. The next step E44B is the calculation of the degradations on the signal S, using the primitives extracted in step E42B. This calculation uses the corresponding coefficients in the cognitive model selected in step E43B. For example, the degradation due to the density of homogeneous zones in the sky regions of the images of a sky-sea documentary will be: - the density of homogeneous zones in the sky regions of the S-signal images corresponding to this sky-sea documentary multiplied by the corresponding weighting coefficient _{l 1} of the cognitive model selected in step E43B.

Finally, step E45B is the integration of the impairments calculated in step E44B into the cognitive model selected in step E43B, giving the estimate of the perceived quality Q of the signal S.

It should be noted that the order of the steps in this last step E4 for estimating the perceived quality of the signal S is not fixed, the step E43B being for example as an alternative implemented after step E42B, and steps E43B, E44B and E45B grouped in a single step. Similarly, in another variant, only steps E2 and E4 are implemented, using a single indexing method, and a single non-reference perceived quality measurement method, which is adapted to the class or content type. determined in step E1. Finally, in this embodiment of the invention, the non-reference measurement method according to the invention is implemented in a single device G, but a distributed implementation is possible. For example, in a variant, the modules G1 and G2 are implemented in a remote indexing server of the device G, which then implements only the modules G3 and G4.

Claims

1. Measurement method without reference to the perceived quality of a signal (S) carrying data representative of a content at least visual or sound, characterized in that it comprises the steps of:

indexing (E2) said content, determining at least one class (C) of semantic membership of said content,

and estimating (E4) the perceived quality of said signal (S) by using a measurement method without reference of perceived quality, said method being adapted to said membership class (C) or a type (t) of content of said class (C).

The non-reference measurement method according to claim 1, wherein said estimating step (E4) is preceded by a step of selecting (E3) said measurement method without reference of perceived quality among a set of measurement methods. without reference adapted to various types or classes of semantic membership of contents.

The non-reference measurement method according to claim 2, wherein said selecting step (E3) takes into account at least one parameter (k) of speed of the measurement method to be selected or a parameter (p) of accuracy of measure of perceived quality.

4. A method of measurement without reference according to any one of claims 1 to 3, wherein said indexing step (E2) is preceded by a step of selecting (E1) said indexing method among a set of methods indexing method, said selecting step (E1) taking into account at least one parameter (k) of speed of the indexing method to be selected or a parameter (p) of measurement accuracy of perceived quality.

The non-reference measurement method according to any one of claims 1 to 4, wherein said indexing step (E2) takes into account at least one meta-information on said signal (S).

The non-reference measurement method according to any one of claims 1 to 5, wherein said estimating step (E4) comprises an extraction sub-step (E41A) in said signal (S), primitives related to estimating the perceived quality of said signal (S), said primitives being adapted to said membership class (C) or content type (t) of said class (C).

The non-reference measurement method according to any one of claims 1 to 5, wherein said signal (S) carries data representative of one or more images, and wherein said estimating step (E4) comprises a sub-step of analyzing (E41 B) at least one of said images, and an extraction sub-step (E42B) of primitives related to the estimation of the perceived quality of said signal (S) as a function of a or more regions identified in said image during the analyzing step (E41 B).

8. Device (G) for measuring without reference to the perceived quality of a signal (S) bearing data representative of at least a visual or sound content, characterized in that it comprises:

means for obtaining indexing data (G2) for said content, determining at least one class (C) of semantic membership of said content,

and means for estimating (G4) the perceived quality of said signal (S) by using a measurement method without reference of perceived quality, said method being adapted to said class (C) of membership or to a type (t ) of content of said class.

9. Terminal incorporating the device (G) according to claim 8.

A computer program comprising instructions for implementing the method of any one of claims 1 to 7 when executed on a computer.