Product evaluation method and system and storage medium

Abstract

The invention provides a product evaluation method and system and a storage medium. The method comprises the following steps: constructing a score matrix according to a product sample image and an expert score; determining key emotion vocabularies in the emotion sample vocabularies according to the scoring matrix, and constructing model training data according to the key emotion vocabularies and the model sample images; performing model training on the product evaluation model according to the model training data until the product evaluation model converges; and inputting a to-be-evaluated product image into the converged product evaluation model for emotion evaluation to obtain an emotion score, and generating a product evaluation result according to the emotion score. According to the embodiment of the invention, model training is carried out on the product evaluation model through the model training data, so that the converged product evaluation model can accurately capture emotion preferences of the user on different product images, emotion evaluation can be effectively carried out on the input to-be-evaluated product image, an emotion score is obtained, and the user experience is improved. And a product evaluation result can be automatically generated based on the emotion score.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a product evaluation method and system and a storage medium. BACKGROUND

[0002] With the wide application of generative artificial intelligence in the field of product design, designers can quickly generate a large number of creative product images through keywords. However, it is still a key challenge for current product research and development to quickly screen products that meet enterprise needs and have high product evaluation from a large number of image schemes.

[0003] In the existing product evaluation process, product evaluation is generally performed in a manner of artificial subjective experience, which leads to strong subjectivity of product evaluation and reduces the accuracy of product evaluation. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a product evaluation method, system and storage medium to solve the problem of low product evaluation accuracy in the prior art.

[0005] The embodiments of the present application are implemented as follows. A product evaluation method comprises: obtaining product sample images and expert scores for the product sample images, and constructing a score matrix according to the product sample images and the expert scores, wherein the expert scores include score values of different emotional sample words; determining key emotional words in the emotional sample words according to the score matrix, and constructing model training data according to the key emotional words and model sample images; performing model training on a product evaluation model according to the model training data until the product evaluation model converges; inputting a product image to be evaluated into the converged product evaluation model for emotional evaluation to obtain an emotional score, and generating a product evaluation result according to the emotional score.

[0006] Preferably, determining the key emotional words in the emotional sample words according to the score matrix comprises: obtaining score values corresponding to the emotional sample words in the score matrix to obtain matrix score values, and calculating squares of the matrix score values to obtain projection intensity values; for the same emotional sample words, calculating sums between corresponding projection intensity values to obtain attribute projection areas, and calculating a sum between the attribute projection areas to obtain an attribute projection total area; calculating a proportion between the attribute projection area and the attribute projection total area to obtain an attribute weight, and determining the key emotional words in the emotional sample words according to the attribute weight.

[0007] Preferably, the model training data is constructed according to the key emotional vocabulary and the model sample image, comprising: A Likert scale is generated according to the key emotional vocabulary and the model sample image, and score information and morphological coding information for the Likert scale are obtained; A score label is generated according to the score information, and a morphological label is generated according to the morphological coding information; The model sample image is labeled according to the score label and the morphological label, and the model training data is obtained.

[0008] Preferably, the product evaluation model is trained according to the model training data, comprising: The model sample image is feature extracted to obtain image features, and the morphological label is feature extracted to obtain morphological features; The image features and the morphological features are calculated by a cross-modal attention mechanism to obtain cross-modal attention features, and the cross-modal attention features are fused with the image features to obtain fused features; The fused features are evaluated for emotional prediction scores, and the emotional prediction scores and the score labels are loss calculated to obtain model loss; The product evaluation model is updated according to the model loss until the product evaluation model converges.

[0009] Preferably, after the emotional score is generated, it further comprises: The preferred product image in the product image to be evaluated is determined according to the emotional score, and the preferred product image is edge expanded to obtain an expanded product image; The user is virtually augmented and displayed according to the expanded product image, and the user's user score is obtained; The target product image in the preferred product image is determined according to the user score.

[0010] Preferably, the score matrix is constructed as: Wherein, X is the score matrix, x nm is the score of the n th product sample image to the m th emotional sample vocabulary; Wherein, S j is the score of the jThe attribute projection area of ​​each sentiment sample word. x ij For the first i The product sample image for the first j Scoring of each sentiment sample word; The formula used to calculate the ratio between the attribute projection area and the total attribute projection area includes: in, n The total number of sample images for the model. m The total number of the sentiment sample words. S stotal The total area projected onto the attribute.

[0011] Preferably, cross-modal attention features are calculated by performing a cross-modal attention mechanism on the image features and the morphological features to obtain cross-modal attention features, including: An image query matrix, an image key matrix, and an image value matrix are generated based on the image features, and an image morphological attention score is calculated based on the image query matrix, the image key matrix, and the image value matrix. A morphological query matrix, a morphological key matrix, and a morphological value matrix are generated based on the morphological features, and a morphological image attention score is calculated based on the morphological query matrix, the morphological key matrix, and the morphological value matrix. The image morphology attention score and the morphology image attention score are normalized to obtain the image morphology attention weight and the morphology image attention weight. The morphology query matrix is ​​weighted according to the image morphology attention weight to obtain image side-modal features, and the image query matrix is ​​weighted according to the morphology image attention weight to obtain morphology side-modal features. The image-side cross-modal features and the morphology-side cross-modal features are subjected to residual fusion to obtain multimodal residual features, and attention mechanisms are applied to the multimodal residual features to obtain the cross-modal attention features.

[0012] Another objective of this invention is to provide a product evaluation system, the system comprising: A matrix construction module is used to acquire product sample images and expert ratings for the product sample images, and to construct a rating matrix based on the product sample images and the expert ratings, wherein the expert ratings include rating values ​​for different emotional sample words; The model training module is used to determine the key sentiment words in the sentiment sample vocabulary based on the scoring matrix, and to construct model training data based on the key sentiment words and model sample images. model training is performed on the product evaluation model according to the model training data until the product evaluation model converges; The product evaluation module is configured to input a product image to be evaluated into the converged product evaluation model to perform sentiment evaluation, obtain a sentiment score, and generate a product evaluation result according to the sentiment score.

[0013] Preferably, the model training module is further configured to: obtain the score values corresponding to the sentiment sample words in the score matrix to obtain matrix score values, and calculate squares of the matrix score values to obtain projection intensity values; For the same sentiment sample words, the sum of the projection intensity values corresponding thereto is calculated to obtain attribute projection areas, and the sum of the attribute projection areas is calculated to obtain an attribute projection total area; The proportion between the attribute projection area and the attribute projection total area is calculated to obtain an attribute weight, and the key sentiment words in the sentiment sample words are determined according to the attribute weight.

[0014] In the embodiment of the present application, the score matrix can be effectively constructed by obtaining product sample images and expert scores, the key sentiment words in the sentiment sample words can be automatically determined based on the score matrix, the subjectivity and isolation of artificial experience evaluation are avoided, an objective quantitative benchmark is provided for the product evaluation model, the product evaluation model is model trained by model training data, the converged product evaluation model can accurately capture the emotional preferences of users for different product images, can effectively perform sentiment evaluation on the input product image to be evaluated to obtain a sentiment score, and can automatically generate a product evaluation result based on the sentiment score. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of the product evaluation method provided by the first embodiment of the present application; Figure 2 is a structural schematic diagram of the product evaluation system provided by the second embodiment of the present application; Figure 3 is a structural schematic diagram of the terminal device provided by the third embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0017] In order to illustrate the technical solutions of the present application, the following specific embodiments are used to illustrate the technical solutions of the present application.

[0018] Embodiment One Referring to Figure 1 is a flowchart of a product evaluation method provided by the first embodiment of the present application. The product evaluation method can be applied to any device or system. The product evaluation method comprises the following steps: In step S10, product sample images and expert scores for the product sample images are obtained, and a score matrix is constructed according to the product sample images and the expert scores. The product sample images can be set according to requirements. Experts and users are invited to perform emotional scoring on the product sample images in different emotional sample vocabulary dimensions, and the average scores of the emotional sample vocabulary corresponding to each product sample image are calculated to obtain the expert scores.

[0019] In this step, the emotional sample vocabulary includes structured data and unstructured data. The structured data sources cover product academic papers and reports, and the unstructured data comes from user evaluation areas in different websites and product emotional corpora, ensuring the diversity and timeliness of the data. The processing flow strictly follows the standard specifications: first, the word frequency of the obtained text is counted, combined with the Chinese stop word table and product field professional stop words, high-frequency emotional words are screened out, then data cleaning is performed, the vocabulary format is unified, and merging operation is completed through a synonym dictionary. To realize the standardization classification of emotional words, a field emotional dictionary and a Word2Vec word vector model are introduced. The former is used to directly match words with clear emotional tendencies, and the latter maps words with similar semantics to a unified category by calculating word vector similarity, for example, “happy”, “happy”, and “joy” are all classified as “happy”. Finally, the K-means algorithm is used to perform clustering analysis on the processed word vectors, and the rationality of the clustering number is verified by the contour coefficient. When the contour coefficient reaches 0.75 or more, the final clustering result is determined, and the emotional sample vocabulary is obtained, ensuring that the words in each category have high semantic homogeneity. For example, for the automotive interior product, the final emotional sample vocabulary includes 8 categories of emotional words such as “comfort”, “safety”, “luxury”, “technology”, “environmental protection”, “warmth”, “atmosphere”, and “elegance”. This provides accurate vocabulary data support for subsequent emotional analysis.

[0020] Optionally, the constructed score matrix is: wherein, X is the score matrix, x nm is the score of the i-th product sample image to the j-th emotional sample vocabulary. n m is the score of the i-th product sample image to the j-th emotional sample vocabulary.

[0021] ​Step S20, determining key emotional words in the emotional sample words according to the score matrix, and constructing model training data according to the key emotional words and model sample images; Wherein, the key emotional words in the emotional sample words can be automatically determined based on the score matrix, avoiding the subjectivity and isolation of manual experience evaluation, and providing an objective quantitative benchmark for the product evaluation model.

[0022] Optionally, determining the key emotional words in the emotional sample words according to the score matrix comprises: Obtaining the score value corresponding to the emotional sample words in the score matrix to obtain a matrix score value, and calculating the square of the matrix score value to obtain a projection intensity value; For the same emotional sample words, the sum of the corresponding projection intensity values is calculated to obtain an attribute projection area, and the sum of the attribute projection areas is calculated to obtain an attribute projection total area; The proportion between the attribute projection area and the attribute projection total area is calculated to obtain an attribute weight, and the key emotional words in the emotional sample words are determined according to the attribute weight; Wherein, for each emotional sample word, the scores in all samples are squared and added to obtain the attribute projection area of the emotional sample word. Taking the emotional sample word "Comfort" as an example, its attribute projection area S1 is the sum of the squares of all scores of the emotional sample word in 127 samples, i.e. (S1=3.982+4.022+42+4.122+…+3.722). Secondly, the attribute projection areas of the remaining emotional sample words are calculated in the same way, and the attribute projection areas of the emotional sample words are added. Through quantitative calculation of the attribute weight, the top three emotional sample words are finally determined as the key emotional words.

[0023] Optionally, the formula for calculating the attribute projection area comprises: Wherein, S j is the attribute projection area of the i-th emotional sample word, j x ij is the score of the i-th product sample image to the i-th emotional sample word. i j

[0024] Further, the formula for calculating the proportion between the attribute projection area and the attribute projection total area comprises: Wherein, n ​​​a total number of the model sample images, m a total number of the emotional sample vocabulary, S stotal a total area of the attribute projection.

[0025] Further, the model training data is constructed according to the key emotional vocabulary and the model sample images, and includes: A Likert scale is generated according to the key emotional vocabulary and the model sample images, and score information and morphological coding information for the Likert scale are obtained; wherein the Likert scale is sent to experts and users for scoring to obtain the score information, image coding of the model sample images is obtained, morphological coding is determined based on the image coding, and morphological coding information is generated based on the morphological coding; Score labels are generated according to the score information, morphological labels are generated according to the morphological coding information, and the model sample images are labeled according to the score labels and the morphological labels to obtain the model training data; wherein the score value in the score information is set as the score label, and the coding value in the morphological coding information is set as the morphological label.

[0026] In step S30, the product evaluation model is trained according to the model training data until the product evaluation model converges. The product evaluation model can adopt a multi-modal cross attention network (Multi-Modality Cross Attention Network, MMCA). The MMCA network is composed of three modules, namely an image feature extraction module, a morphological feature extraction module, and a feature fusion module. The image feature extraction module adopts a pre-trained ResNet-18 model. The ResNet-18 network extracts a 512-dimensional feature vector after convolution and pooling operations. The morphological feature extraction module (full connection layer): The morphological feature extraction module processes the input morphological features through two full connection layers (fc1 and fc2). Assuming that the input morphological features contain 6 morphological features, the first layer full connection fc1 maps the input 6-dimensional features to 512-dimensional features.

[0027] Optionally, the model training of the product evaluation model according to the model training data includes: The model sample images are subjected to feature extraction to obtain image features, and the morphological labels are subjected to feature extraction to obtain morphological features; wherein the model sample images are subjected to convolution and pooling processing by the image feature extraction module to obtain the image features, and the morphological labels are subjected to full connection processing by the morphological feature extraction module to obtain the morphological features. The image feature and the morphology feature are subjected to cross-modal attention mechanism calculation to obtain a cross-modal attention feature, and the cross-modal attention feature is fused with the image feature to obtain a fused feature; in order to effectively fuse the image feature and the morphology feature, a cross-modal attention mechanism is adopted. The image feature and the morphology feature vector are input into the attention mechanism, in order to enable the attention mechanism to process these features, it is necessary to convert them into a suitable shape by increasing the batch dimension, and then use a multi-head attention mechanism to calculate the attention output; The fused feature is subjected to sentiment evaluation to obtain a sentiment prediction score, and the sentiment prediction score and the score label are subjected to loss calculation to obtain a model loss; in order to further generate a final prediction result, the image feature and the attention output are spliced, and a final prediction is performed through a fully connected layer. The image feature and the attention mechanism output are spliced to obtain a feature vector with a length of 1024, and the spliced feature is mapped to an output space through a fully connected layer to predict a sentiment score index and obtain a sentiment prediction score; The product evaluation model is subjected to parameter updating according to the model loss until the product evaluation model converges; in which, mean square error (MSE) is used as a loss function to optimize the model, the MSE calculates the square difference between the predicted value and the true label, and the loss is minimized through a gradient descent method, and in the model test stage, root mean square error and determination coefficient are used as two evaluation indexes to measure the prediction performance of the model.

[0028] Further, the image feature and the morphology feature are subjected to cross-modal attention mechanism calculation to obtain a cross-modal attention feature, including: An image query matrix, an image key matrix and an image value matrix are generated according to the image feature, and an image morphology attention score is calculated according to the image query matrix, the image key matrix and the image value matrix; A morphology query matrix, a morphology key matrix and a morphology value matrix are generated according to the morphology feature, and a morphology image attention score is calculated according to the morphology query matrix, the morphology key matrix and the morphology value matrix; The image morphology attention score and the morphology image attention score are subjected to normalization processing to obtain an image morphology attention weight and a morphology image attention weight; The morphology query matrix is weighted according to the image morphology attention weight to obtain an image side cross-modal feature, and the image query matrix is weighted according to the morphology image attention weight to obtain a morphology side cross-modal feature; The image side cross-modal feature and the morphology side cross-modal feature are subjected to residual fusion to obtain a multi-modal residual feature, and the multi-modal residual feature is subjected to attention mechanism calculation to obtain the cross-modal attention feature.

[0029] In step S40, the product image to be evaluated is input into the converged product evaluation model for sentiment evaluation, to obtain a sentiment score, and a product evaluation result is generated based on the sentiment score. The product image to be evaluated is effectively evaluated in terms of sentiment by inputting it into the converged product evaluation model, and a sentiment score is obtained. Based on the sentiment score, a product evaluation result can be automatically generated. The greater the sentiment score, the higher the sentiment score of the product image to be evaluated corresponding to the sentiment score.

[0030] Optionally, after generating the product evaluation result based on the sentiment score, the method further includes: The preferred product image in the product image to be evaluated is determined based on the sentiment score, and the preferred product image is edge-extended to obtain an extended product image. In the sentiment score, the top 15 product images to be evaluated are determined as the preferred product image. By edge-extending the preferred product image, a 2D image can be effectively converted into a panoramic image, presenting details such as spatial layout and material texture, and preparing high-fidelity visual materials for virtual augmented display. The user is provided with virtual augmented display based on the extended product image, a user score of the user is obtained, and a target product image in the preferred product image is determined based on the user score.

[0031] In this embodiment, by obtaining product sample images and expert scores, a score matrix can be effectively constructed. Based on the score matrix, key sentiment words in the sentiment sample vocabulary can be automatically determined, avoiding the subjectivity and isolation of manual experience evaluation. An objective quantitative benchmark is provided for the product evaluation model. The product evaluation model is trained by model training data, so that the converged product evaluation model can accurately capture the emotional preferences of users for different product images, effectively evaluate the input product image to be evaluated in terms of sentiment, and automatically generate a product evaluation result based on the sentiment score.

[0032] In this embodiment, by collecting user sentiment words and calculating weights, Top 3 is selected as the core evaluation index (key sentiment words). Secondly, the MMCA model is trained to fuse images, shape codes, and sentiment scores, to establish a mapping relationship between user sentiment and products, and to realize quantitative prediction of unknown schemes. Finally, an immersive experience environment is constructed through virtual augmented display, and the best scheme (target product image) is finally determined based on user feedback.

[0033] In this embodiment, evaluation indicators for product evaluation were constructed. Mathematical modeling preserved the inherent relationships between indicators, avoiding the subjectivity and isolation of traditional indicator selection, and providing a scientific and objective quantitative benchmark for the evaluation system. A multimodal evaluation model based on MMCA was developed and innovatively applied to the design evaluation field. The model achieves deep integration through a three-layer architecture. This not only accurately captures user emotional preferences but also enhances the objectivity and comprehensiveness of the evaluation. A dual-channel evaluation framework integrating MMCA and virtual augmented display was established, improving the efficiency of screening massive amounts of solutions and bridging the gap between static evaluation and actual user experience, making the evaluation results more aligned with real needs and providing full-process support for the final optimization of product design solutions.

[0034] Example 2 Please see Figure 2 This is a schematic diagram of the structure of the product evaluation system 100 provided in the second embodiment of the present invention, including: The matrix construction module 10 is used to acquire product sample images and expert scores for the product sample images, and to construct a scoring matrix based on the product sample images and the expert scores, wherein the expert scores include score values ​​for different emotional sample words.

[0035] Optionally, the constructed rating matrix is ​​as follows: in, X For the rating matrix, x nm For the first n The product sample image for the first m The rating of each emotional sample word.

[0036] The model training module 11 is used to determine the key sentiment words in the sentiment sample vocabulary according to the scoring matrix, and to construct model training data based on the key sentiment words and model sample images.

[0037] The product evaluation model is trained using the model training data until the product evaluation model converges.

[0038] Optionally, the model training module 11 is further configured to: obtain the rating value corresponding to the sentiment sample word in the rating matrix, obtain the matrix rating value, and calculate the square of the matrix rating value to obtain the projection intensity value; For the same emotional sample word, the sum of the corresponding projection intensity values ​​is calculated to obtain the attribute projection area, and the sum of the attribute projection areas is calculated to obtain the total attribute projection area. The proportion between the attribute projection area and the total attribute projection area is calculated to obtain an attribute weight, and the key emotional vocabulary in the emotional sample vocabulary is determined according to the attribute weight.

[0039] Preferably, the formula used to calculate the attribute projection area includes: wherein, S j is the attribute projection area of the i-th emotional sample vocabulary, j ij is the score of the i-th product sample image to the i-th emotional sample vocabulary; x ij is the score of the i-th product sample image to the i-th emotional sample vocabulary; i j The formula used to calculate the proportion between the attribute projection area and the total attribute projection area includes: wherein, is the total number of the model sample images, n is the total number of the emotional sample vocabulary, m stotal is the total attribute projection area. S Further, the model training module 11 is further configured to generate a Likert scale according to the key emotional vocabulary and the model sample image, and obtain scoring information and morphological coding information for the Likert scale;

[0040] generate a scoring label according to the scoring information, and generate a morphological label according to the morphological coding information; label mark the model sample image according to the scoring label and the morphological label, to obtain the model training data. Further, the model training module 11 is further configured to perform feature extraction on the model sample image to obtain image features, and perform feature extraction on the morphological label to obtain morphological features;

[0041] perform cross-modal attention mechanism calculation on the image features and the morphological features to obtain cross-modal attention features, and fuse the cross-modal attention features with the image features to obtain fused features; perform emotion evaluation on the fused features to obtain an emotional prediction score, and perform loss calculation on the emotional prediction score and the scoring label to obtain a model loss; update the parameters of the product evaluation model according to the model loss until the product evaluation model converges.

[0042] ​Preferably, the model training module 11 is further configured to generate an image query matrix, an image key matrix and an image value matrix according to the image features, and calculate an image modality attention score according to the image query matrix, the image key matrix and the image value matrix; generate a modality query matrix, a modality key matrix and a modality value matrix according to the modality features, and calculate a modality image attention score according to the modality query matrix, the modality key matrix and the modality value matrix; normalize the image modality attention score and the modality image attention score to obtain an image modality attention weight and a modality image attention weight; weight the modality query matrix according to the image modality attention weight to obtain an image side cross-modal feature, and weight the image query matrix according to the modality image attention weight to obtain a modality side cross-modal feature; fuse the image side cross-modal feature and the modality side cross-modal feature by residual to obtain a multi-modal residual feature, and calculate the cross-modal attention feature by attention mechanism on the multi-modal residual feature.

[0043] The product evaluation module 12 is configured to input a product image to be evaluated into the converged product evaluation model to perform sentiment evaluation, obtain a sentiment score, and generate a product evaluation result according to the sentiment score.

[0044] Optionally, the product evaluation module 12 is further configured to determine a preferred product image in the product image to be evaluated according to the sentiment score, and perform edge expansion on the preferred product image to obtain an expanded product image; virtually enhance and display the user according to the expanded product image, and obtain a user score of the user; determine a target product image in the preferred product image according to the user score.

[0045] In this embodiment, by obtaining product sample images and expert scores, a score matrix can be effectively constructed, key sentiment words in the sentiment sample words can be automatically determined based on the score matrix, subjectivity and isolation of manual experience evaluation are avoided, an objective quantitative benchmark is provided for the product evaluation model, the product evaluation model is trained by model training data, the converged product evaluation model can accurately capture the emotional preferences of users for different product images, can effectively perform sentiment evaluation on the input product image to be evaluated to obtain a sentiment score, and can automatically generate a product evaluation result based on the sentiment score.

[0046] Embodiment Three Figure 3 is a structural block diagram of a terminal device 2 provided by the third embodiment of the present application. As shown in Figure 3As shown, the terminal device 2 of the embodiment includes a processor 20, a memory 21, and a computer program 22, such as a program of product evaluation method, stored in the memory 21 and executable on the processor 20. The processor 20 implements the steps in each of the embodiments of the product evaluation method when executing the computer program 22.

[0047] For example, the computer program 22 can be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program 22 in the terminal device 2. The terminal device can include, but is not limited to, the processor 20 and the memory 21.

[0048] The processor 20 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0049] The memory 21 can be an internal storage unit of the terminal device 2, such as a hard disk or a memory of the terminal device 2. The memory 21 can also be an external storage device of the terminal device 2, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 21 can include both the internal storage unit and the external storage device of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 can also be used to temporarily store data that has been output or will be output.

[0050] In addition, each functional module in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0051] If the integrated module is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. The computer readable storage medium can be non-volatile or volatile. Based on this understanding, all or part of the processes in the above-mentioned embodiments can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of each method embodiment described above can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable storage medium can include any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content included in the computer readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable storage medium does not include electric carrier signal and telecommunication signal.

[0052] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A product evaluation method, characterized in that, The method includes: Obtain product sample images and expert ratings for the product sample images, and construct a rating matrix based on the product sample images and the expert ratings, wherein the expert ratings include rating values ​​for different emotional sample words; The key sentiment words in the sentiment sample vocabulary are determined based on the scoring matrix, and model training data is constructed based on the key sentiment words and model sample images. The product evaluation model is trained based on the model training data until the product evaluation model converges. The product image to be evaluated is input into the converged product evaluation model to perform sentiment evaluation, obtain a sentiment score, and generate a product evaluation result based on the sentiment score.

2. The product evaluation method as described in claim 1, characterized in that, Based on the scoring matrix, the key sentiment words in the sentiment sample vocabulary are determined, including: Obtain the rating value corresponding to the emotional sample words in the rating matrix to obtain the matrix rating value, and calculate the square of the matrix rating value to obtain the projection intensity value; For the same emotional sample word, the sum of the corresponding projection intensity values ​​is calculated to obtain the attribute projection area, and the sum of the attribute projection areas is calculated to obtain the total attribute projection area. The ratio between the attribute projection area and the total attribute projection area is calculated to obtain the attribute weight, and the key sentiment words in the sentiment sample vocabulary are determined based on the attribute weight.

3. The product evaluation method as described in claim 1, characterized in that, Model training data is constructed based on the key sentiment terms and model sample images, including: A Likert scale is generated based on the key sentiment words and the model sample images, and rating information and morphological encoding information for the Likert scale are obtained. A rating label is generated based on the rating information, and a morphological label is generated based on the morphological encoding information; The model sample images are labeled according to the rating labels and the morphological labels to obtain the model training data.

4. The product evaluation method as described in claim 3, characterized in that, The product evaluation model is trained based on the model training data, including: Feature extraction is performed on the model sample image to obtain image features, and feature extraction is performed on the morphological label to obtain morphological features; Cross-modal attention mechanisms are used to calculate the image features and the morphological features to obtain cross-modal attention features, and the cross-modal attention features are fused with the image features to obtain fused features; Sentiment evaluation is performed on the fused features to obtain a sentiment prediction score, and loss is calculated on the sentiment prediction score and the score label to obtain the model loss; The parameters of the product evaluation model are updated based on the model loss until the product evaluation model converges.

5. The product evaluation method as described in claim 1, characterized in that, After generating product evaluation results based on the aforementioned sentiment score, the following is also included: Based on the sentiment score, a preferred product image is determined from the product images to be evaluated, and the edge of the preferred product image is expanded to obtain an expanded product image; The user is virtually enhanced based on the extended product image, and the user's rating is obtained. The target product image in the preferred product images is determined based on the user rating.

6. The product evaluation method as described in claim 2, characterized in that, The constructed rating matrix is ​​as follows: in, X For the rating matrix, x nm For the first n The product sample image for the first m Scoring of each sentiment sample word; in, S j For the first j The attribute projection area of ​​each sentiment sample word. x ij For the first i The product sample image for the first j Scoring of each sentiment sample word; The formula used to calculate the ratio between the attribute projection area and the total attribute projection area includes: in, n The total number of sample images for the model. m The total number of the sentiment sample words. S stotal The total area projected onto the attribute.

7. The product evaluation method as described in claim 4, characterized in that, Cross-modal attention features are obtained by performing cross-modal attention mechanism calculations on the image features and the morphological features, including: An image query matrix, an image key matrix, and an image value matrix are generated based on the image features, and an image morphological attention score is calculated based on the image query matrix, the image key matrix, and the image value matrix. A morphological query matrix, a morphological key matrix, and a morphological value matrix are generated based on the morphological features, and a morphological image attention score is calculated based on the morphological query matrix, the morphological key matrix, and the morphological value matrix. The image morphology attention score and the morphology image attention score are normalized to obtain the image morphology attention weight and the morphology image attention weight. The morphology query matrix is ​​weighted according to the image morphology attention weight to obtain image side-modal features, and the image query matrix is ​​weighted according to the morphology image attention weight to obtain morphology side-modal features. The image-side cross-modal features and the morphology-side cross-modal features are subjected to residual fusion to obtain multimodal residual features, and attention mechanisms are applied to the multimodal residual features to obtain the cross-modal attention features.

8. A product evaluation system, characterized in that, The system includes: A matrix construction module is used to acquire product sample images and expert ratings for the product sample images, and to construct a rating matrix based on the product sample images and the expert ratings, wherein the expert ratings include rating values ​​for different emotional sample words; The model training module is used to determine the key sentiment words in the sentiment sample vocabulary based on the scoring matrix, and to construct model training data based on the key sentiment words and model sample images. The product evaluation model is trained based on the model training data until the product evaluation model converges. The product evaluation module is used to input the image of the product to be evaluated into the converged product evaluation model to perform sentiment evaluation, obtain a sentiment score, and generate a product evaluation result based on the sentiment score.

9. The product evaluation system as described in claim 8, characterized in that, The model training module is also used for: Obtain the rating value corresponding to the emotional sample words in the rating matrix to obtain the matrix rating value, and calculate the square of the matrix rating value to obtain the projection intensity value; For the same emotional sample word, the sum of the corresponding projection intensity values ​​is calculated to obtain the attribute projection area, and the sum of the attribute projection areas is calculated to obtain the total attribute projection area. The ratio between the attribute projection area and the total attribute projection area is calculated to obtain the attribute weight, and the key sentiment words in the sentiment sample vocabulary are determined based on the attribute weight.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images