A sound quality assessment method based on multi-dimensional fusion analysis

Abstract

The present invention relates to the technical field of sound quality assessment, and discloses a sound quality assessment method of multi-dimensional fusion analysis, the method comprising: extracting semantic features of a video to be dubbed to obtain video semantic features; fusing the video semantic features and the video text semantic features to obtain video fusion features, and performing sentiment analysis based on the video fusion features to obtain video sentiment classification results; performing multi-feature audio sentiment analysis on the video dubbing, comparing the video dubbing sentiment recognition results with the video sentiment classification results, and obtaining sound quality assessment results. The present invention extracts multi-scale features of spatiotemporal association on video frame images, extracts contextual subtitle features and full-text features on video subtitles, fuses image features and subtitle text features, obtains multi-dimensional fusion video features, compares video dubbing emotions with video emotions, and the comparison result is the consistency of the dubbing emotions with the video emotions, thereby realizing multi-dimensional fusion sound quality assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of sound quality assessment, and in particular to a sound quality assessment method of multi-dimensional fusion analysis. Background Art

[0002] With the rapid development of science and technology, audio technology is increasingly used in various fields, including music production, movie sound effects, speech recognition, virtual reality, etc. As an important part of audio technology, sound quality evaluation directly affects the user's auditory experience and the market competitiveness of products. In the process of video dubbing, traditional sound quality evaluation methods mainly rely on subjective listening tests and objective physical measurements, and require manual verification of the degree of match with the emotions of the video content. However, with the increasing complexity and diversity of audio and video content, a single evaluation method can no longer fully reflect the true situation of sound quality. How to accurately and efficiently evaluate the sound quality of video dubbing has become an important issue that needs to be solved urgently. Summary of the invention

[0003] In view of this, the present invention provides a sound quality assessment method of multi-dimensional fusion analysis, which deeply integrates scene analysis and takes into account multiple factors, such as frequency response, dynamic range, distortion, etc., to achieve comprehensive and accurate sound quality assessment results.

[0004] To achieve the above object, the present invention provides a sound quality assessment method of multi-dimensional fusion analysis, comprising the following steps:

[0005] S1: extracting semantic features from the video to be dubbed to obtain video semantic features, wherein multi-scale features associated with time and space are fused as an implementation method of the feature extraction;

[0006] S2: extracting text semantic features of the dubbed video subtitles in combination with the context scene to obtain the video text semantic features;

[0007] S3: fusing the video semantic features and the video text semantic features to obtain video fusion features, and performing sentiment analysis based on the video fusion features to obtain video sentiment classification results, wherein the multimodal fusion method is an implementation method of the feature fusion;

[0008] S4: Perform multi-feature audio sentiment analysis on the video dubbing to obtain the video dubbing sentiment recognition result;

[0009] S5: Compare the video dubbing emotion recognition result with the video emotion classification result. If the two are the same, the dubbing sound quality passes, otherwise it fails.

[0010] As a further improvement method of the present invention:

[0011] Optionally, the step S1 includes extracting semantic features from the video to be dubbed, including:

[0012] Get the video to be dubbed, where the video consists of continuous video frame images. The obtained video is:

[0013] ;

[0014] in:

[0015] represents the acquired video;

[0016] Video N consecutive video frame images in, Video The nth video frame image in;

[0017] About Video Semantic feature extraction is performed to obtain video semantic features, wherein the spatiotemporal-correlated multi-scale feature fusion is the implementation method of the feature extraction, and the video semantic feature extraction process is as follows:

[0018] S11: Video Any video frame image in is segmented into image elements, where the video frame image The image element segmentation result is:

[0019] ;

[0020] in:

[0021] Represents a video frame image The image element in the Ath row and the Bth column, A represents the number of rows of the image element segmentation of the video frame image, and B represents the number of columns of the image element segmentation of the video frame image;

[0022] The pixel matrix representation of any image element is:

[0023] ;

[0024] in:

[0025] Represents image element The pixel value of the pixel at row x and column y in , , X represents the number of pixel rows of any image element, and Y represents the number of pixel columns of any image element;

[0026] S12: Extract image features of image elements, where image elements The image features are:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035]

[0036] in:

[0037] Represents image element Image features;

[0038] The image element Texture features, color features and contour features;

[0039] The grayscale values ​​are And two horizontally adjacent pixels in the image element The number of groups that appear in The gray value of the pixel in row x and column y is ;

[0040] Image element The color value of the pixel in the xth row and yth column in the R, G, and B color channels;

[0041] Represents image element The chromaticity value of the pixel at the xth row and the yth column;

[0042] Represents image element The gradient value of the pixel at row x and column y in , Represents image element The gradient change ratio of the pixel in the xth row and yth column;

[0043] S13: Scale fusion is performed on the multi-level image features of the image element in a spatiotemporal correlation manner to form scale fusion features of the image element, where the image element The scale fusion feature is ;

[0044] S14: splicing the scale fusion features of all image elements in the video frame image to form the semantic features of the video frame image, where the video frame image The semantic features are:

[0045] ;

[0046] ;

[0047] ;

[0048] in:

[0049] Represents a video frame image The semantic features of

[0050] Represents the length of the scale fusion feature of the image element;

[0051] represents the video semantic extraction matrix, * represents the convolution operator;

[0052] Representation scale fusion feature The semantic weight of

[0053] represents an exponential function with a natural constant as base;

[0054] Indicates Line The position weights of the column image elements;

[0055] Composition video Video semantic features:

[0056] ;

[0057] in:

[0058] Video The semantic features of the video.

[0059] Optionally, in step S13, the scale fusion of multi-level image features of image elements is performed in a spatiotemporal association manner, including:

[0060] S131: Extracting spatially associated texture information from texture features, where texture features The spatially associated texture information is:

[0061] ;

[0062] ;

[0063] in:

[0064] Representing texture features In the embodiment of the present invention, Represents the texture energy information of pixels at different positions, Represents texture related information of pixels at different positions;

[0065] S132: Extracting spatially associated color information from color features, where color features Spatially associated color information for:

[0066] ;

[0067] in:

[0068] Representation Matrix The mean of the elements in , Representation Matrix The standard deviation of the elements in ;

[0069] S133: Combine temporal association to generate scale fusion features, where image elements Scale fusion features for:

[0070] ;

[0071] in:

[0072] represents the temporal convolution matrix, Representing contour features valid information.

[0073] Optionally, in step S2, extracting text semantic features of the video subtitles to be dubbed in combination with the context scene includes:

[0074] Get the video subtitles to be dubbed, where the video subtitles are a subtitle sequence consisting of the subtitles corresponding to each video frame image. The obtained video subtitles are in the form of:

[0075] ;

[0076] in:

[0077] Indicates the acquired video subtitles;

[0078] Indicates the N subtitles corresponding to N consecutive video frame images, Represents a video frame image corresponding subtitles;

[0079] Treating Video Subtitles for Dubbing The video text semantic feature extraction is performed in combination with the context scene, wherein the video text semantic feature extraction process is as follows:

[0080] S21: Perform word vectorization on N subtitles to obtain a word vector sequence corresponding to each subtitle, where The corresponding word vector sequence is ; In the embodiment of the present invention, the model selected for word vectorization representation is the Word2vec model;

[0081] S22: Subtitles for Video Perform full-text feature extraction, where the full-text feature extraction process is:

[0082] S221: setting the current subtitle feature to be extracted as the context subtitle feature of the nth subtitle, where the initial value of n is 1;

[0083] S222: Extract the context subtitle features of the nth subtitle, where the feature extraction formula is:

[0084] ;

[0085] in:

[0086] represents the subtitle feature extraction matrix;

[0087] represents the contextual subtitle feature of the nth subtitle; in the embodiment of the present invention, represents a unit vector whose elements are all 1;

[0088] S223: Let n=n+1 until , get the full text feature that integrates N contextual subtitle features ;

[0089] S23: Based on full-text features Semantic features are extracted from the word vector sequence of the subtitle to obtain the semantic features of the subtitle, where the word vector sequence The semantic feature extraction formula is:

[0090] ;

[0091] ;

[0092] in:

[0093] T stands for transpose;

[0094] Representing word vector sequence The e-th word vector in , Representing word vector sequence The total number of word vectors in ;

[0095] Indicates the preset subtitle length threshold;

[0096] Representing word vector sequence The corresponding nth subtitle Semantic features of subtitles;

[0097] S24: Constructing video text semantic features .

[0098] Optionally, in step S3, the video semantic feature and the video text semantic feature are fused to obtain a video fusion feature, including:

[0099] The video semantic features and video text semantic features The video fusion features are obtained by fusion, and the fusion process is as follows:

[0100] S31: Video semantic features and video text semantic features The semantic features in are nonlinearly mapped respectively, where the semantic features and The nonlinear mapping formula is:

[0101] ;

[0102] in:

[0103] Semantic features and The nonlinear characteristics of

[0104] They are video semantic features and video text semantic features The mapping matrix of semantic features in ;

[0105] S32: Calculate the outer product of two nonlinear features corresponding to the same video frame image, where the nonlinear feature The outer product is:

[0106] ;

[0107] in:

[0108] Represents nonlinear characteristics The outer product of

[0109] represents the outer product operator;

[0110] S33: Global pooling of the outer product, where the outer product The global pooling processing formula is:

[0111] ;

[0112] in:

[0113] Represents the outer product The global pooling result of

[0114] S34: Normalize the global pooling processing result to obtain the fusion features corresponding to the video frame image, and form a video frame image with the fusion features of the continuous video frame images. Video fusion features :

[0115] ;

[0116] ;

[0117] in:

[0118] express The normalized result is represents the L1 norm.

[0119] Optionally, performing sentiment analysis according to the video fusion features to obtain a video sentiment classification result includes:

[0120] Based on the video fusion features Perform sentiment analysis to obtain the video sentiment classification results, where the sentiment analysis process is as follows:

[0121] Obtain the emotion classification matrix representing the emotion states of different videos, where the emotion states of videos include 8 categories, namely anger, disgust, fear, sadness, expectation, happiness, surprise and trust;

[0122] Calculate the video fusion features The video emotional state corresponding to each fusion feature in The corresponding video emotional state is:

[0123] ;

[0124] in:

[0125] ;

[0126] Represents the emotion classification matrix corresponding to the emotion state of the u-th video;

[0127] represents the L2 norm;

[0128] Represents fusion features And the fusion features corresponding to the previous video frame image Correction parameters for the emotional state of the u-th video;

[0129] Represents fusion features The corresponding video emotional state, where , corresponding to 8 types of video emotional states;

[0130] Construct the video sentiment classification results: .

[0131] Optionally, performing multi-feature audio sentiment analysis on the video dubbing in step S4 includes:

[0132] Get the video dubbing, where the video dubbing is a subtitle audio sequence corresponding to N subtitles. The obtained video dubbing is in the form of:

[0133] ;

[0134] in:

[0135] Indicates the acquired video dubbing;

[0136] Indicates the N subtitle audios corresponding to N consecutive subtitles. Subtitles Corresponding subtitle audio;

[0137] Dubbing videos Perform multi-feature audio sentiment analysis to obtain the video dubbing sentiment recognition result. The multi-feature audio sentiment analysis process is as follows:

[0138] S41: Dubbing Video The subtitle audio in is discretized to obtain the discretized subtitle audio signal, where the subtitle audio Corresponding subtitle audio signal for , Indicates subtitle audio signal The signal value at the mth audio signal sampling time in , where M represents the total number of audio signal sampling times of the subtitle audio;

[0139] S42: Pre-emphasize the subtitle audio signal, wherein the subtitle audio signal The corresponding pre-emphasis signal for:

[0140] ;

[0141] ;

[0142] in:

[0143] is the signal value The pre-emphasis result is Represents the pre-emphasis coefficient. Set to 0.95;

[0144] S43: Calculate and obtain the spectrum representation of the pre-emphasized signal, where the pre-emphasized signal The spectrum of is expressed as:

[0145] ;

[0146] ;

[0147] in:

[0148] Indicates the pre-emphasis signal The spectrum representation of Indicates the pre-emphasis signal In the spectrum of Fourier transform point number k, q represents the imaginary unit;

[0149] S44: Using the power spectrum corresponding to the spectrum representation as the input of the Mel filter bank to obtain the Mel cepstrum coefficients corresponding to the pre-emphasized signal, wherein the pre-emphasized signal The corresponding Mel cepstral coefficient is ;

[0150] S45: Extract features of the pre-emphasized signal, the spectrum representation, and the Mel-frequency cepstrum coefficients to obtain multi-level audio features, where the pre-emphasized signal The corresponding feature extraction result is , spectrum representation The corresponding feature extraction result is , Mel cepstral coefficient The corresponding feature extraction result is ;

[0151] S46: performing fusion processing on the multi-level audio features to form a multi-level audio fusion feature of the subtitle audio signal, wherein the subtitle audio signal Corresponding multi-level audio fusion features for:

[0152] ;

[0153] in:

[0154] represents the softmax function;

[0155] Represents the direct sum operator of vectors;

[0156] S47: Generate an emotion recognition result corresponding to the subtitle audio signal to form a video dubbing emotion recognition result s:

[0157] ;

[0158] ;

[0159] in:

[0160] Indicates subtitle audio signal The corresponding emotion recognition results, , corresponding to 8 categories of video emotional states.

[0161] Optionally, comparing the video dubbing emotion recognition result with the video emotion classification result in step S5 includes:

[0162] Compare the video dubbing emotion recognition result s with the video emotion classification result S. If the two are the same, the dubbing sound quality passes, otherwise it fails. The comparison process of the video emotion classification results is as follows:

[0163] S51: Obtain the video emotion state of the video dubbing emotion recognition result s and the video emotion classification result S in the same video frame image, and obtain a video emotion state comparison group of N video frame images, wherein the video emotion state comparison group of the nth video frame image is ;

[0164] S52: Calculate and obtain the emotional state comparison result of each video emotional state comparison group, where the video emotional state comparison group The comparison results of emotional states are:

[0165] ;

[0166] ;

[0167] in:

[0168] Indicates the comparison group of video emotional states Comparison results of emotional states;

[0169] represents the intermediate parameter, Represents the window adjustment coefficient;

[0170] H represents the size of the emotional condition comparison window;

[0171] S53: Calculate and compare the video dubbing emotion recognition result s and the video emotion classification result S :

[0172] ;

[0173] like If it is smaller than the preset comparison threshold, it means that the video dubbing emotion recognition result s is the same as the video emotion classification result S.

[0174] Optionally, in the step S45, feature extraction is performed on the pre-emphasized signal, the spectrum representation and the Mel-frequency cepstral coefficients respectively to obtain multi-level audio features, including:

[0175] S451: Pre-emphasis signal The signal values ​​are clustered to obtain K clusters, each of which has several signal values. The mean of all signal values ​​in the cluster is used as the cluster number, and the cluster number is used to replace the signal value in the cluster to form a quantitative feature composed of the cluster number. ;

[0176] S452: Spectrum representation Perform multi-layer convolution processing and use the convolution result of the last layer as ;

[0177] S453: Using Bidirectional Long Short-Term Memory Network to Mel Cepstral Coefficients Extract features and construct .

[0178] In order to solve the above problem, the present invention provides an electronic device, the electronic device comprising:

[0179] A memory storing at least one instruction;

[0180] Communication interface, enabling electronic equipment to communicate; and

[0181] The processor executes the instructions stored in the memory to implement the sound quality assessment method of multi-dimensional fusion analysis described above.

[0182] In order to solve the above problems, the present invention also provides a computer-readable storage medium, in which at least one instruction is stored. The at least one instruction is executed by a processor in an electronic device to implement the sound quality assessment method of multi-dimensional fusion analysis described above.

[0183] Compared with the prior art, the present invention proposes a sound quality assessment method based on multi-dimensional fusion analysis, which has the following advantages:

[0184] First, this scheme proposes a video semantic feature extraction method. By collecting video frame images and video subtitles of the video to be dubbed, multi-scale feature extraction with temporal and spatial correlation is performed on the video frame images to obtain scale features that characterize the texture features, color features and contour features of the video frame images. The scale feature spatial information of the adjacent image elements and the temporal change information of the scale features are combined to perform spatiotemporal correlation to form video semantic features. Contextual subtitle features and full-text feature extraction are performed on the video subtitles. According to the full-text feature extraction results, semantic feature extraction is performed on each subtitle in combination with word vector position information, attention weight and full-text feature correction. The video semantic features and video text semantic features in the same video scene are fused to obtain fused features that fuse the video frame image semantics and the subtitle text semantics. Multi-dimensional sentiment analysis combining image dimension and text dimension is realized to obtain the video emotion corresponding to each video frame.

[0185] At the same time, this scheme proposes an audio emotion analysis and emotion comparison method. By obtaining the video dubbing and performing multi-scale decomposition processing on the video dubbing, a pre-emphasized signal representing the change of the audio signal, a spectrum representation representing the frequency domain information, and a Mel-frequency cepstral coefficient representing the energy change are obtained. Feature extraction is performed on the scale decomposition results to obtain multi-level audio features of the video dubbing, and emotion analysis is performed to obtain the video dubbing emotion recognition results and video emotion classification results. The dubbing sound quality is evaluated according to the degree of matching between the two. The higher the matching degree, the higher the consistency between the dubbing emotion and the video emotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0186] Figure 1 A schematic diagram of a flow chart of a sound quality assessment method of multi-dimensional fusion analysis provided by an embodiment of the present invention;

[0187] Figure 2 A schematic diagram of the structure of an electronic device for implementing a sound quality assessment method for multi-dimensional fusion analysis provided by an embodiment of the present invention.

[0188] In the figure: 1 electronic device, 10 processor, 11 memory, 12 program, 13 communication interface.

[0189] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0190] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0191] The embodiment of the present application provides a method for evaluating sound quality by multi-dimensional fusion analysis. The execution subject of the method for evaluating sound quality by multi-dimensional fusion analysis includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the method for evaluating sound quality by multi-dimensional fusion analysis can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.

[0192] Embodiment 1:

[0193] S1: Extract semantic features from the video to be dubbed to obtain video semantic features.

[0194] In the step S1, semantic features of the video to be dubbed are extracted, including:

[0195] Get the video to be dubbed, where the video consists of continuous video frame images. The obtained video is:

[0196] ;

[0197] in:

[0198] represents the acquired video;

[0199] Video N consecutive video frame images in, Video The nth video frame image in;

[0200] About Video Semantic feature extraction is performed to obtain video semantic features, wherein the spatiotemporal-correlated multi-scale feature fusion is the implementation method of the feature extraction, and the video semantic feature extraction process is as follows:

[0201] S11: Video Any video frame image in is segmented into image elements, where the video frame image The image element segmentation result is:

[0202] ;

[0203] in:

[0204] Represents a video frame image The image element in the Ath row and the Bth column, A represents the number of rows of the image element segmentation of the video frame image, and B represents the number of columns of the image element segmentation of the video frame image;

[0205] The pixel matrix representation of any image element is:

[0206] ;

[0207] in:

[0208] Represents image element The pixel value of the pixel at row x and column y in , , X represents the number of pixel rows of any image element, and Y represents the number of pixel columns of any image element;

[0209] S12: Extract image features of image elements, where image elements The image features are:

[0210] ;

[0211] ;

[0212] ;

[0213] ;

[0214] ;

[0215] ;

[0216] ;

[0217] ;

[0218]

[0219] in:

[0220] Represents image element Image features;

[0221] The image element Texture features, color features and contour features;

[0222] The grayscale values ​​are And two horizontally adjacent pixels in the image element The number of groups that appear in The gray value of the pixel in row x and column y is ;

[0223] Image element The color value of the pixel in the xth row and yth column in the R, G, and B color channels;

[0224] Represents image element The chromaticity value of the pixel at the xth row and the yth column;

[0225] Represents image element The gradient value of the pixel at row x and column y in , Represents image element The gradient change ratio of the pixel in the xth row and yth column;

[0226] S13: Scale fusion is performed on the multi-level image features of the image element in a spatiotemporal correlation manner to form scale fusion features of the image element, where the image element The scale fusion feature is ;

[0227] S14: splicing the scale fusion features of all image elements in the video frame image to form the semantic features of the video frame image, where the video frame image The semantic features are:

[0228] ;

[0229] ;

[0230] ;

[0231] in:

[0232] Represents a video frame image The semantic features of

[0233] Represents the length of the scale fusion feature of the image element;

[0234] represents the video semantic extraction matrix, * represents the convolution operator;

[0235] Representation scale fusion feature The semantic weight of

[0236] represents an exponential function with a natural constant as base;

[0237] Indicates Line The position weights of the column image elements;

[0238] Composition video Video semantic features:

[0239] ;

[0240] in:

[0241] Video The semantic features of the video.

[0242] In the step S13, the multi-level image features of the image element are scale-fused by adopting a spatiotemporal association method, including:

[0243] S131: Extracting spatially associated texture information from texture features, where texture features The spatially associated texture information is:

[0244] ;

[0245] ;

[0246] in:

[0247] Representing texture features In the embodiment of the present invention, Represents the texture energy information of pixels at different positions, Represents texture related information of pixels at different positions;

[0248] S132: Extracting spatially associated color information from color features, where color features Spatially associated color information for:

[0249] ;

[0250] in:

[0251] Representation Matrix The mean of the elements in , Representation Matrix The standard deviation of the elements in ;

[0252] S133: Combine temporal association to generate scale fusion features, where image elements Scale fusion features for:

[0253] ;

[0254] in:

[0255] represents the temporal convolution matrix, Representing contour features valid information.

[0256] S2: Extract text semantic features of the video subtitles to be dubbed in combination with the context scene to obtain video text semantic features.

[0257] In the step S2, text semantic features of the video subtitles to be dubbed are extracted in combination with the context scene, including:

[0258] Get the video subtitles to be dubbed, where the video subtitles are a subtitle sequence consisting of the subtitles corresponding to each video frame image. The obtained video subtitles are in the form of:

[0259] ;

[0260] in:

[0261] Indicates the acquired video subtitles;

[0262] Indicates the N subtitles corresponding to N consecutive video frame images, Represents a video frame image corresponding subtitles;

[0263] Treating Video Subtitles for Dubbing The video text semantic feature extraction is performed in combination with the context scene, wherein the video text semantic feature extraction process is as follows:

[0264] S21: Perform word vectorization on N subtitles to obtain a word vector sequence corresponding to each subtitle, where The corresponding word vector sequence is ; In the embodiment of the present invention, the model selected for word vectorization representation is the Word2vec model;

[0265] S22: Subtitles for Video Perform full-text feature extraction, where the full-text feature extraction process is:

[0266] S221: setting the current subtitle feature to be extracted as the context subtitle feature of the nth subtitle, where the initial value of n is 1;

[0267] S222: Extract the context subtitle features of the nth subtitle, where the feature extraction formula is:

[0268] ;

[0269] in:

[0270] represents the subtitle feature extraction matrix;

[0271] represents the contextual subtitle feature of the nth subtitle; in the embodiment of the present invention, represents a unit vector whose elements are all 1;

[0272] S223: Let n=n+1 until , get the full text feature that integrates N contextual subtitle features ;

[0273] S23: Based on full-text features Semantic features are extracted from the word vector sequence of the subtitle to obtain the semantic features of the subtitle, where the word vector sequence The semantic feature extraction formula is:

[0274] ;

[0275] ;

[0276] in:

[0277] T stands for transpose;

[0278] Representing word vector sequence The e-th word vector in , Representing word vector sequence The total number of word vectors in ;

[0279] Indicates the preset subtitle length threshold;

[0280] Representing word vector sequence The corresponding nth subtitle Semantic features of subtitles;

[0281] S24: Constructing video text semantic features .

[0282] S3: The video semantic features and the video text semantic features are fused to obtain the video fusion features, and sentiment analysis is performed based on the video fusion features to obtain the video sentiment classification results.

[0283] In the step S3, the video semantic features and the video text semantic features are fused to obtain the video fusion features, including:

[0284] The video semantic features and video text semantic features The video fusion features are obtained by fusion, and the fusion process is as follows:

[0285] S31: Video semantic features and video text semantic features The semantic features in are nonlinearly mapped respectively, where the semantic features and The nonlinear mapping formula is:

[0286] ;

[0287] in:

[0288] Semantic features and The nonlinear characteristics of

[0289] They are video semantic features and video text semantic features The mapping matrix of semantic features in ;

[0290] S32: Calculate the outer product of two nonlinear features corresponding to the same video frame image, where the nonlinear feature The outer product is:

[0291] ;

[0292] in:

[0293] Represents nonlinear characteristics The outer product of

[0294] represents the outer product operator;

[0295] S33: Global pooling of the outer product, where the outer product The global pooling processing formula is:

[0296] ;

[0297] in:

[0298] Represents the outer product The global pooling result of

[0299] S34: Normalize the global pooling processing result to obtain the fusion features corresponding to the video frame image, and form a video frame image with the fusion features of the continuous video frame images. Video fusion features :

[0300] ;

[0301] ;

[0302] in:

[0303] express The normalized result is represents the L1 norm.

[0304] The sentiment analysis is performed according to the video fusion features to obtain the video sentiment classification result, including:

[0305] Based on the video fusion features Perform sentiment analysis to obtain the video sentiment classification results, where the sentiment analysis process is as follows:

[0306] Obtain the emotion classification matrix representing the emotion states of different videos, where the emotion states of videos include 8 categories, namely anger, disgust, fear, sadness, expectation, happiness, surprise and trust;

[0307] Calculate the video fusion features The video emotional state corresponding to each fusion feature in The corresponding video emotional state is:

[0308] ;

[0309] in:

[0310] ;

[0311] Represents the emotion classification matrix corresponding to the emotion state of the u-th video;

[0312] represents the L2 norm;

[0313] Represents fusion features And the fusion features corresponding to the previous video frame image Correction parameters for the emotional state of the u-th video;

[0314] Represents fusion features The corresponding video emotional state, where , corresponding to 8 types of video emotional states;

[0315] Construct the video sentiment classification results: .

[0316] S4: Perform multi-feature audio sentiment analysis on the video dubbing to obtain the video dubbing sentiment recognition result.

[0317] The step S4 performs multi-feature audio sentiment analysis on the video dubbing, including:

[0318] Get the video dubbing, where the video dubbing is a subtitle audio sequence corresponding to N subtitles. The obtained video dubbing is in the form of:

[0319] ;

[0320] in:

[0321] Indicates the acquired video dubbing;

[0322] Indicates the N subtitle audios corresponding to N consecutive subtitles. Subtitles Corresponding subtitle audio;

[0323] Dubbing videos Perform multi-feature audio sentiment analysis to obtain the video dubbing sentiment recognition result. The multi-feature audio sentiment analysis process is as follows:

[0324] S41: Dubbing Video The subtitle audio in is discretized to obtain the discretized subtitle audio signal, where the subtitle audio Corresponding subtitle audio signal for , Indicates subtitle audio signal The signal value at the mth audio signal sampling time in , where M represents the total number of audio signal sampling times of the subtitle audio;

[0325] S42: Pre-emphasize the subtitle audio signal, wherein the subtitle audio signal The corresponding pre-emphasis signal for:

[0326] ;

[0327] ;

[0328] in:

[0329] is the signal value The pre-emphasis result is Represents the pre-emphasis coefficient. Set to 0.95;

[0330] S43: Calculate and obtain the spectrum representation of the pre-emphasized signal, where the pre-emphasized signal The spectrum of is expressed as:

[0331] ;

[0332] ;

[0333] in:

[0334] Indicates the pre-emphasis signal The spectrum representation of Indicates the pre-emphasis signal In the spectrum of Fourier transform point number k, q represents the imaginary unit;

[0335] S44: Using the power spectrum corresponding to the spectrum representation as the input of the Mel filter bank to obtain the Mel cepstrum coefficients corresponding to the pre-emphasized signal, wherein the pre-emphasized signal The corresponding Mel cepstral coefficient is ;

[0336] S45: Extract features of the pre-emphasized signal, the spectrum representation, and the Mel-frequency cepstrum coefficients to obtain multi-level audio features, where the pre-emphasized signal The corresponding feature extraction result is , spectrum representation The corresponding feature extraction result is , Mel cepstral coefficient The corresponding feature extraction result is ;

[0337] S46: performing fusion processing on the multi-level audio features to form a multi-level audio fusion feature of the subtitle audio signal, wherein the subtitle audio signal Corresponding multi-level audio fusion features for:

[0338] ;

[0339] in:

[0340] represents the softmax function;

[0341] Represents the direct sum operator of vectors;

[0342] S47: Generate an emotion recognition result corresponding to the subtitle audio signal to form a video dubbing emotion recognition result s:

[0343] ;

[0344] ;

[0345] in:

[0346] Indicates subtitle audio signal The corresponding emotion recognition results, , corresponding to 8 categories of video emotional states.

[0347] S5: Compare the video dubbing emotion recognition result with the video emotion classification result. If the two are the same, the dubbing sound quality passes, otherwise it fails.

[0348] The step S5 compares the video dubbing emotion recognition result with the video emotion classification result, including:

[0349] Compare the video dubbing emotion recognition result s with the video emotion classification result S. If the two are the same, the dubbing sound quality passes, otherwise it fails. The comparison process of the video emotion classification results is as follows:

[0350] S51: Obtain the video emotion state of the video dubbing emotion recognition result s and the video emotion classification result S in the same video frame image, and obtain a video emotion state comparison group of N video frame images, wherein the video emotion state comparison group of the nth video frame image is ;

[0351] S52: Calculate and obtain the emotional state comparison result of each video emotional state comparison group, where the video emotional state comparison group The comparison results of emotional states are:

[0352] ;

[0353] ;

[0354] in:

[0355] Indicates the comparison group of video emotional states Comparison results of emotional states;

[0356] represents the intermediate parameter, Represents the window adjustment coefficient;

[0357] H represents the size of the emotional condition comparison window;

[0358] S53: Calculate and compare the video dubbing emotion recognition result s and the video emotion classification result S :

[0359] ;

[0360] like If it is smaller than the preset comparison threshold, it means that the video dubbing emotion recognition result s is the same as the video emotion classification result S.

[0361] Embodiment 2:

[0362] like Figure 2 , which is a schematic diagram of the structure of an electronic device for implementing a sound quality assessment method of multi-dimensional fusion analysis provided by an embodiment of the present invention.

[0363] The electronic device 1 may include a processor 10 , a memory 11 , a communication interface 13 and a bus, and may also include a computer program stored in the memory 11 and executable on the processor 10 , such as a program 12 .

[0364] The memory 11 includes at least one type of readable storage medium, and the readable storage medium includes a flash memory, a mobile hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory, etc.), a magnetic memory, a disk, an optical disk, etc. In some embodiments, the memory 11 may be an internal storage unit of the electronic device 1, such as a mobile hard disk of the electronic device 1. In other embodiments, the memory 11 may also be an external storage device of the electronic device 1, such as a plug-in mobile hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 1. Further, the memory 11 may also include both an internal storage unit of the electronic device 1 and an external storage device. The memory 11 may not only be used to store application software and various types of data installed in the electronic device 1, such as the code of the program 12, etc., but also be used to temporarily store data that has been output or is to be output.

[0365] The processor 10 may be composed of an integrated circuit in some embodiments, for example, a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and a combination of various control chips, etc. The processor 10 is the control core (Control Unit) of the electronic device, and uses various interfaces and lines to connect various components of the entire electronic device, and executes or executes programs or modules stored in the memory 11 (such as the sound quality assessment program 12 for realizing multi-dimensional fusion analysis), and calls data stored in the memory 11 to execute various functions of the electronic device 1 and process data.

[0366] The communication interface 13 may include a wired interface and / or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), which is generally used to establish a communication connection between the electronic device 1 and other electronic devices, and to achieve connection and communication between internal components of the electronic device.

[0367] The bus may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. The bus is configured to realize connection and communication between the memory 11 and at least one processor 10, etc.

[0368] Figure 2 Only an electronic device with components is shown, and those skilled in the art will understand that Figure 2 The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown in the figure, or combine certain components, or arrange the components differently.

[0369] For example, although not shown, the electronic device 1 may also include a power source (such as a battery) for supplying power to various components. Preferably, the power source may be logically connected to the at least one processor 10 through a power management device, so that the power management device can realize functions such as charging management, discharging management, and power consumption management. The power source may also include one or more DC or AC power sources, recharging devices, power failure detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 1 may also include a variety of sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be repeated here.

[0370] Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), or a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, and an OLED (Organic Light-Emitting Diode) touch device. The display may also be appropriately referred to as a display screen or a display unit, which is used to display information processed in the electronic device 1 and to display a visual user interface.

[0371] It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.

[0372] The program 12 stored in the memory 11 in the electronic device 1 is a combination of multiple instructions. When running in the processor 10, it can achieve:

[0373] Extracting semantic features of the video to be dubbed to obtain video semantic features;

[0374] Extract text semantic features of the video subtitles to be dubbed in combination with the context scene to obtain the video text semantic features;

[0375] The video semantic features and the video text semantic features are fused to obtain the video fusion features, and the sentiment analysis is performed based on the video fusion features to obtain the video sentiment classification results;

[0376] Perform multi-feature audio sentiment analysis on video dubbing to obtain the video dubbing sentiment recognition result;

[0377] Compare the video dubbing emotion recognition result with the video emotion classification result. If the two are the same, the dubbing sound quality passes, otherwise it fails.

[0378] Specifically, the specific implementation method of the processor 10 for the above instructions can refer to Figure 1 to Figure 2 The description of the relevant steps in the corresponding embodiments will not be repeated here.

[0379] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0380] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0381] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A sound quality assessment method based on multi-dimensional fusion analysis, characterized in that: The method comprises: S1: Extract semantic features of the video to be dubbed to obtain video semantic features; S2: extracting text semantic features of the dubbed video subtitles in combination with the context scene to obtain the video text semantic features; S3: Fusing the video semantic features and the video text semantic features to obtain the video fusion features, and performing sentiment analysis based on the video fusion features to obtain the video sentiment classification results; S4: Perform multi-feature audio sentiment analysis on the video dubbing to obtain the video dubbing sentiment recognition result; S5: Compare the video dubbing emotion recognition result with the video emotion classification result. If the two are the same, the dubbing sound quality passes, otherwise it fails; In the step S1, semantic features of the video to be dubbed are extracted, including: Get the video to be dubbed, where the video consists of continuous video frame images. The obtained video is: ； in: represents the acquired video; Video N consecutive video frame images in, Video The nth video frame image in; About Video Semantic features are extracted to obtain video semantic features, where the process of extracting video semantic features is as follows: S11: Video Any video frame image in is segmented into image elements, where the video frame image The image element segmentation result is: ； in: Represents a video frame image The image element in the Ath row and the Bth column, A represents the number of rows of the image element segmentation of the video frame image, and B represents the number of columns of the image element segmentation of the video frame image; The pixel matrix representation of any image element is: ； in: Represents image element The pixel value of the pixel at row x and column y in , , X represents the number of pixel rows of any image element, and Y represents the number of pixel columns of any image element; S12: Extract image features of image elements, where image elements The image features are: ； ； ； ； ； ； ； in: Represents image element Image features; The image element Texture features, color features and contour features; The grayscale values ​​are And two horizontally adjacent pixels in the image element The number of groups that appear in The gray value of the pixel in row x and column y is ; Image element The color value of the pixel in the xth row and yth column in the R, G, and B color channels; Represents image element The chromaticity value of the pixel at the xth row and the yth column; Represents image element The gradient value of the pixel at row x and column y in , Represents image element The gradient change ratio of the pixel in the xth row and yth column; S13: Scale fusion is performed on the multi-level image features of the image element in a spatiotemporal correlation manner to form scale fusion features of the image element, where the image element The scale fusion feature is ; S14: splicing the scale fusion features of all image elements in the video frame image to form the semantic features of the video frame image, where the video frame image The semantic features are: ； ； ； in: Represents a video frame image The semantic features of Represents the length of the scale fusion feature of the image element; represents the video semantic extraction matrix, * represents the convolution operator; Representation scale fusion feature The semantic weight of represents an exponential function with a natural constant as base; Indicates Line The position weights of the columns are: Composition video Video semantic features: ； in: Video The semantic features of the video.

2. A sound quality assessment method of multi-dimensional fusion analysis as claimed in claim 1, characterized in that: In the step S13, the multi-level image features of the image element are scale-fused by adopting a spatiotemporal association method, including: S131: Extracting spatially associated texture information from texture features, where texture features The spatially associated texture information is: ； ； in: Representing texture features Spatially associated texture information; S132: Extracting spatially associated color information from color features, where color features Spatially associated color information for: ； in: Representation Matrix The mean of the elements in , Representation Matrix The standard deviation of the elements in ; S133: Combine temporal association to generate scale fusion features, where image elements Scale fusion features for: ； in: represents the temporal convolution matrix, Representing contour features valid information.

3. The sound quality assessment method of multi-dimensional fusion analysis as claimed in claim 1, characterized in that: In the step S2, text semantic features of the video subtitles to be dubbed are extracted in combination with the context scene, including: Get the video subtitles to be dubbed, where the video subtitles are a subtitle sequence consisting of the subtitles corresponding to each video frame image. The obtained video subtitles are in the form of: ； in: Indicates the acquired video subtitles; Indicates the N subtitles corresponding to N consecutive video frame images, Represents a video frame image corresponding subtitles; Treating Video Subtitles for Dubbing The video text semantic feature extraction is performed in combination with the context scene, wherein the video text semantic feature extraction process is as follows: S21: Perform word vectorization on N subtitles to obtain a word vector sequence corresponding to each subtitle, where The corresponding word vector sequence is ; S22: Subtitles for Video Perform full-text feature extraction to obtain full-text features ; S23: Based on full-text features Semantic features are extracted from the word vector sequence of the subtitle to obtain the semantic features of the subtitle, where the word vector sequence The semantic feature extraction formula is: ； ； in: T stands for transpose; Representing word vector sequence The e-th word vector in , Representing word vector sequence The total number of word vectors in ; Indicates the preset subtitle length threshold; Representing word vector sequence The corresponding nth subtitle Semantic features of subtitles; S24: Constructing video text semantic features .

4. A sound quality assessment method of multi-dimensional fusion analysis as claimed in claim 3, characterized in that: In the step S3, the video semantic features and the video text semantic features are fused to obtain the video fusion features, including: The video semantic features and video text semantic features The video fusion features are obtained by fusion, and the fusion process is as follows: S31: Video semantic features and video text semantic features The semantic features in are nonlinearly mapped respectively, where the semantic features and The nonlinear mapping formula is: ； in: Semantic features and The nonlinear characteristics of They are video semantic features and video text semantic features The mapping matrix of semantic features in ; S32: Calculate the outer product of two nonlinear features corresponding to the same video frame image, where the nonlinear feature The outer product is: ； in: Represents nonlinear characteristics The outer product of represents the outer product operator; S33: Global pooling of the outer product, where the outer product The global pooling processing formula is: ； in: Represents the outer product The global pooling result of S34: Normalize the global pooling processing result to obtain the fusion features corresponding to the video frame image, and form a video frame image with the fusion features of the continuous video frame images. Video fusion features : ； ； in: express The normalized result is represents the L1 norm; Sentiment analysis is performed based on video fusion features to obtain video sentiment classification results.

5. A sound quality assessment method of multi-dimensional fusion analysis as claimed in claim 4, characterized in that: The sentiment analysis is performed according to the video fusion features to obtain the video sentiment classification result, including: Based on the video fusion features Perform sentiment analysis to obtain the video sentiment classification results, where the sentiment analysis process is as follows: Obtain the emotion classification matrix representing the emotion states of different videos, where the emotion states of videos include 8 categories, namely anger, disgust, fear, sadness, expectation, happiness, surprise and trust; Calculate the video fusion features The video emotional state corresponding to each fusion feature in The corresponding video emotional state is: ； in: ； Represents the emotion classification matrix corresponding to the emotion state of the u-th video; represents the L2 norm; Represents fusion features And the fusion features corresponding to the previous video frame image Correction parameters for the emotional state of the u-th video; Represents fusion features The corresponding video emotional state, where , corresponding to 8 types of video emotional states; Construct the video sentiment classification results: .

6. The sound quality assessment method of multi-dimensional fusion analysis as claimed in claim 1, characterized in that: The step S4 performs multi-feature audio sentiment analysis on the video dubbing, including: Get the video dubbing, where the video dubbing is a subtitle audio sequence corresponding to N subtitles. The obtained video dubbing is in the form of: ； in: V represents the acquired video dubbing; Indicates the N subtitle audios corresponding to N consecutive subtitles. Subtitles Corresponding subtitle audio; Dubbing videos Perform multi-feature audio sentiment analysis to obtain the video dubbing sentiment recognition result. The multi-feature audio sentiment analysis process is as follows: S41: Dubbing Video The subtitle audio in is discretized to obtain the discretized subtitle audio signal, where the subtitle audio Corresponding subtitle audio signal for , Indicates subtitle audio signal The signal value at the mth audio signal sampling time in , where M represents the total number of audio signal sampling times of the subtitle audio; S42: Pre-emphasize the subtitle audio signal, wherein the subtitle audio signal The corresponding pre-emphasis signal for: ； ； in: is the signal value The pre-emphasis result is Represents the pre-emphasis coefficient. Set to 0.95; S43: Calculate and obtain the spectrum representation of the pre-emphasized signal, where the pre-emphasized signal The spectrum of is expressed as: ； ； in: Indicates the pre-emphasis signal The spectrum representation of Indicates the pre-emphasis signal In the spectrum of Fourier transform point number k, q represents the imaginary unit; S44: Using the power spectrum corresponding to the spectrum representation as the input of the Mel filter bank to obtain the Mel cepstrum coefficients corresponding to the pre-emphasized signal, wherein the pre-emphasized signal The corresponding Mel cepstral coefficient is ; S45: Extract features of the pre-emphasized signal, the spectrum representation, and the Mel-frequency cepstrum coefficients to obtain multi-level audio features, where the pre-emphasized signal The corresponding feature extraction result is , spectrum representation The corresponding feature extraction result is , Mel cepstral coefficient The corresponding feature extraction result is ; S46: performing fusion processing on the multi-level audio features to form a multi-level audio fusion feature of the subtitle audio signal, wherein the subtitle audio signal Corresponding multi-level audio fusion features for: ； in: represents the softmax function; Represents the direct sum operator of vectors; S47: Generate an emotion recognition result corresponding to the subtitle audio signal to form a video dubbing emotion recognition result s: ； ； in: Indicates subtitle audio signal The corresponding emotion recognition results, , corresponding to 8 categories of video emotional states.

7. A method for evaluating sound quality by multi-dimensional fusion analysis as claimed in claim 6, characterized in that: The step S5 compares the video dubbing emotion recognition result with the video emotion classification result, including: Compare the video dubbing emotion recognition result s with the video emotion classification result S. If the two are the same, the dubbing sound quality passes, otherwise it fails. The comparison process of the video emotion classification results is as follows: S51: Obtain the video emotion state of the video dubbing emotion recognition result s and the video emotion classification result S in the same video frame image, and obtain a video emotion state comparison group of N video frame images, wherein the video emotion state comparison group of the nth video frame image is ; S52: Calculate and obtain the emotional state comparison result of each video emotional state comparison group, where the video emotional state comparison group The comparison results of emotional states are: ； ； in: Indicates the comparison group of video emotional states Comparison results of emotional states; represents the intermediate parameter, Represents the window adjustment coefficient; H represents the size of the emotional condition comparison window; S53: Calculate and compare the video dubbing emotion recognition result s and the video emotion classification result S : ； like If it is smaller than the preset comparison threshold, it means that the video dubbing emotion recognition result s is the same as the video emotion classification result S.

8. The sound quality assessment method of multi-dimensional fusion analysis as claimed in claim 6, characterized in that: In the step S45, feature extraction is performed on the pre-emphasized signal, the spectrum representation and the Mel-frequency cepstral coefficients to obtain multi-level audio features, including: S451: Pre-emphasis signal The signal values ​​in are clustered to obtain K clusters, each of which has several signal values. The mean of all signal values ​​in the cluster is used as the cluster number, and the cluster number is used to replace the signal value in the cluster to form a quantitative feature composed of the cluster number. ; S452: Spectrum representation Perform multi-layer convolution processing and use the convolution result of the last layer as ; S453: Using Bidirectional Long Short-Term Memory Network to Mel Cepstral Coefficients Extract features and construct .

Images