Product attribute prediction method, system, device and storage medium

Abstract

The present invention discloses a prediction method, system, device and storage medium for product attributes. The prediction method includes obtaining first image data of a first product to be trained and obtaining a target training set; using a residual network to extract attribute features of the target training set to obtain a first attribute prediction vector and a first feature vector; using a pooling method to perform pooling on the first feature vector to obtain a second feature vector; using a linear classifier and a nonlinear classifier to classify the second feature vector to obtain an intermediate attribute prediction vector, and then establishing a target attribute prediction model for predicting product attributes; obtaining second image data of a target product to be predicted, and predicting the target attribute of the target product through the target attribute prediction model. The present invention overcomes the defects of serious overfitting and high prediction errors in existing prediction methods, improves the prediction accuracy of product attributes, and the entire method has good robustness and practical application capabilities.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, system, device and storage medium for predicting product attributes. Background Art

[0002] At present, intelligent product attribute classification is one of the problems that Internet platforms urgently need to solve. Faced with tens of billions of products on Internet platforms, it is no longer advisable to manually review product attribute classification through Internet servers, and it is even more unrealistic to rely on sellers to fill in product attributes truthfully with a strong sense of self-discipline, because in practice, it is found that most sellers will obtain higher search attention by over-describing products in product titles and piling up false attributes that do not match the products, among which clothing products are particularly prominent.

[0003] Existing product attribute prediction methods mainly include: keyword and rule-based methods, natural language processing-based methods, image classification-based methods, and end-to-end translation methods based on image-generated titles. However, these prediction methods have the following defects: (1) For keyword and rule-based methods, constructing keyword dictionaries and observing matching patterns requires experience and is extremely time-consuming and labor-intensive, and it is impossible to discover new words, new matching patterns, etc.; (2) For natural language processing-based methods, the processing effect is not ideal for short sentences with a large number of word stacks; (3) For image classification-based methods, clothing attributes such as collar type are fine-grained classification problems in local areas, and an attention positioning mechanism is required to prevent the overfitting phenomenon of inferring collars from skirts; (4) For end-to-end translation methods based on image-generated titles, it is difficult to translate clothing images into titles, which can easily lead to overfitting of the entire system; and unreliable title content will inevitably lead to a high error rate in the prediction of the training model. Summary of the invention

[0004] The technical problem to be solved by the present invention is that the method for determining product attributes in the prior art has defects such as serious overfitting and high prediction error rate. The purpose is to provide a method, system, device and storage medium for predicting product attributes.

[0005] The present invention solves the above technical problems through the following technical solutions:

[0006] The present invention provides a method for predicting product attributes, the method comprising:

[0007] Acquire first image data of a first product to be trained, and acquire a target training set according to the first image data;

[0008] Performing attribute feature extraction processing on the target training set using a residual network, obtaining a first attribute prediction vector output by the residual network, and obtaining a first feature vector corresponding to a residual module obtained by the residual network during the attribute feature extraction process;

[0009] Performing pooling processing on the first feature vector by using a pooling method to obtain a second feature vector;

[0010] Using a linear classifier and a nonlinear classifier to classify the second feature vector to obtain an intermediate attribute prediction vector;

[0011] The first attribute prediction vector and the intermediate attribute prediction vector are concatenated to establish a target attribute prediction model for predicting product attributes;

[0012] Acquire second image data of a target product to be predicted;

[0013] The second image data is input into the target attribute prediction model to predict the target attribute of the target product.

[0014] Preferably, the step of using a linear classifier and a nonlinear classifier to classify the second feature vector to obtain an intermediate attribute prediction vector includes:

[0015] Using a linear classifier to classify the second feature vector to obtain a second attribute prediction vector;

[0016] Using a nonlinear classifier to classify the second feature vector to obtain a third attribute prediction vector;

[0017] The step of concatenating the first attribute prediction vector and the intermediate attribute prediction vector to establish a target attribute prediction model for predicting product attributes comprises:

[0018] The first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector are concatenated to establish the target attribute prediction model for predicting product attributes.

[0019] Preferably, after the step of performing pooling processing on the first feature vector by using a pooling method to obtain a second feature vector, and before the step of performing classification processing on the second feature vector by using a linear classifier and a nonlinear classifier to obtain an intermediate attribute prediction vector, the step further includes:

[0020] The principal component analysis technology is used to perform dimensionality reduction processing on the second eigenvector to obtain the second eigenvector after dimensionality reduction processing.

[0021] Preferably, the step of acquiring a target training set according to the first image data comprises:

[0022] The first image data is divided into M sub-data sets, and k sub-data sets are selected to form the target training set, where k is less than M and both M and k are positive integers.

[0023] Preferably, the step of acquiring a target training set according to the first image data comprises:

[0024] Divide the first image data into M sub-datasets, and select k of the sub-datasets to form a first training set, where k < M and M and k are both positive integers;

[0025] The first training set is preprocessed by using at least one of color conversion processing, flip processing and cropping processing to obtain the preprocessed target training set.

[0026] Preferably, when the first training set is trimmed, data of a plurality of sub-images corresponding to the first image data are obtained;

[0027] The data of each sub-image constitutes a target training set.

[0028] Preferably, when the first training set is trimmed, data of a plurality of sub-images corresponding to the first image data are obtained;

[0029] The step of using a residual network to extract attribute features from the target training set includes:

[0030] Using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image;

[0031] The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process;

[0032] The step of obtaining a first feature vector corresponding to the residual module obtained by the residual network in the attribute feature extraction process includes:

[0033] Obtaining the first feature vector corresponding to the residual module obtained by the residual network in a process of extracting attribute features at different levels of data of each sub-image;

[0034] The step of performing pooling processing on the first feature vector by using the pooling method to obtain the second feature vector comprises:

[0035] The pooling method is used to perform pooling processing on the first feature vectors corresponding to the data of each sub-image at different levels, so as to obtain the second feature vectors corresponding to the data of each sub-image at different levels.

[0036] Preferably, after using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image, the step of obtaining the first attribute prediction vector includes:

[0037] Obtaining a first sub-prediction vector corresponding to the data of each sub-image at a different level;

[0038] Respectively perform average aggregation processing on the first sub-prediction vectors corresponding to the data of the same sub-image at different levels to obtain the first attribute prediction vector corresponding to the data of each sub-image;

[0039] The step of using a linear classifier to classify the second feature vector to obtain a second attribute prediction vector comprises:

[0040] Using the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels;

[0041] Respectively perform average aggregation processing on the second sub-prediction vectors corresponding to the data of the same sub-image at different levels to obtain the second attribute prediction vector corresponding to the data of each sub-image;

[0042] The step of using a nonlinear classifier to classify the second feature vector to obtain a third attribute prediction vector comprises:

[0043] Using the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the third sub-prediction vectors corresponding to the data of each sub-image at different levels;

[0044] Respectively perform average aggregation processing on the third sub-prediction vectors corresponding to the data of the same sub-image at different levels to obtain the third attribute prediction vector corresponding to the data of each sub-image;

[0045] The step of concatenating the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector to establish the target attribute prediction model for predicting product attributes includes:

[0046] Performing splicing processing on the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector to obtain an intermediate prediction model corresponding to the data of each of the sub-images;

[0047] The intermediate prediction models corresponding to the data of each sub-image are spliced ​​to establish the target attribute prediction model for predicting product attributes.

[0048] Preferably, the step of performing splicing processing on the intermediate prediction models corresponding to the data of each sub-image to establish the target attribute prediction model for predicting product attributes comprises:

[0049] In the case where the intermediate prediction models corresponding to the data of each sub-image have equal weights, the intermediate prediction models corresponding to the data of each sub-image are spliced ​​to establish the target attribute prediction model for predicting product attributes.

[0050] Preferably, after using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image, the step of obtaining the first attribute prediction vector includes:

[0051] Obtaining a first sub-prediction vector corresponding to the data of each sub-image at a different level;

[0052] Using a regression algorithm to perform regression processing on the first sub-prediction vectors corresponding to the data of different sub-images at different levels, respectively, to obtain the first attribute prediction vector corresponding to the first image data;

[0053] The step of using a linear classifier to classify the second feature vector to obtain a second attribute prediction vector comprises:

[0054] Using the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels;

[0055] Using the regression algorithm to perform regression processing on the second sub-prediction vectors corresponding to the data of different sub-images at different levels, respectively, to obtain the second attribute prediction vector corresponding to the first image data;

[0056] The step of using a nonlinear classifier to classify the second feature vector to obtain a third attribute prediction vector comprises:

[0057] Using the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the third sub-prediction vectors corresponding to the data of each sub-image at different levels;

[0058] The regression algorithm is used to perform regression processing on the third sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the third attribute prediction vector corresponding to the first image data.

[0059] Preferably, the step of performing attribute feature extraction processing on the target training set using a residual network to obtain a first attribute prediction vector output by the residual network comprises:

[0060] Using a residual network to perform convolution processing on the target training set to obtain N residual modules, where N is a positive integer;

[0061] Performing pooling processing on the Nth residual module using the pooling method to obtain an intermediate feature vector;

[0062] The dimension of the intermediate feature vector is higher than the dimension of the first feature vector corresponding to the residual module;

[0063] A regression algorithm is used to perform regression processing on the intermediate feature vector to obtain the first attribute prediction vector.

[0064] Preferably, when the residual network includes N residual modules during the attribute feature extraction process, the first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network during the attribute feature extraction process.

[0065] Preferably, the residual network comprises a deep residual network or Inception (a learning model); and / or,

[0066] The linear classifier comprises a support vector machine; and / or,

[0067] The nonlinear classifier includes a regression classification tree or a random forest algorithm; and / or,

[0068] The pooling method includes global average pooling; and / or,

[0069] The regression algorithm includes a multi-class logistic regression algorithm.

[0070] Preferably, the step of inputting the second image data into the target attribute prediction model to predict the target attribute of the target product comprises:

[0071] Inputting the second image data into the target attribute prediction model, and selecting the maximum vector value in the prediction vector corresponding to the target attribute prediction model;

[0072] The attribute corresponding to the maximum vector value is determined as the target attribute of the target product.

[0073] Preferably, the step of dividing the first image data into M sub-datasets and selecting k sub-datasets to form a first training set further includes:

[0074] Select Mk sub-datasets to form a validation set;

[0075] The step of obtaining the target attribute prediction model further includes:

[0076] Obtaining actual attributes of the first product being trained;

[0077] Inputting the sub-dataset in the validation set into the target attribute prediction model to obtain a prediction result of the first product;

[0078] Determine whether the prediction result is consistent with the actual attribute of the trained product. If so, determine that the target attribute prediction model is feasible; otherwise, retrain the target attribute prediction model until the prediction result is consistent with the actual attribute of the first product being trained.

[0079] The present invention also provides a product attribute prediction system, the prediction system further comprising a first image data acquisition module, a target training set acquisition module, a first prediction vector acquisition module, a first feature vector acquisition module, a second feature vector acquisition module, an intermediate prediction vector acquisition module, a prediction model establishment module, a second image data acquisition module and a target attribute prediction module;

[0080] The first image data acquisition module is used to acquire first image data of a first product to be trained;

[0081] The target training set acquisition module is used to acquire a target training set according to the first image data;

[0082] The first prediction vector acquisition module is used to use a residual network to perform attribute feature extraction processing on the target training set to obtain a first attribute prediction vector output by the residual network;

[0083] The first feature vector acquisition module is used to obtain a first feature vector corresponding to the residual module obtained by the residual network during the attribute feature extraction process;

[0084] The second feature vector acquisition module is used to perform pooling processing on the first feature vector by using a pooling method to obtain a second feature vector;

[0085] The intermediate prediction vector acquisition module is used to classify the second feature vector using a linear classifier and a nonlinear classifier to obtain an intermediate attribute prediction vector;

[0086] The prediction model building module is used to perform splicing processing on the first attribute prediction vector and the intermediate attribute prediction vector to establish a target attribute prediction model for predicting product attributes;

[0087] The second image data acquisition module is used to acquire second image data of the target product to be predicted;

[0088] The target attribute prediction module is used to input the second image data into the target attribute prediction model to predict the target attribute of the target product.

[0089] Preferably, the intermediate prediction vector acquisition module is used to classify the second feature vector using a linear classifier to acquire a second attribute prediction vector;

[0090] The intermediate prediction vector acquisition module is further used to classify the second feature vector using a nonlinear classifier to obtain a third attribute prediction vector;

[0091] The prediction model building module is used to perform splicing processing on the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector to establish the target attribute prediction model for predicting product attributes.

[0092] Preferably, the prediction system further comprises a principal component analysis module;

[0093] The principal component analysis module is used to perform dimensionality reduction processing on the second eigenvector using the principal component analysis technology to obtain the second eigenvector after dimensionality reduction processing.

[0094] Preferably, the target training set acquisition module is used to divide the first image data into M sub-data sets, and select k sub-data sets to form the target training set, k<M and M and k are both positive integers.

[0095] Preferably, the target training set acquisition module includes a training set acquisition unit and a preprocessing unit;

[0096] The training set acquisition unit is used to divide the first image data into M sub-data sets, and select k sub-data sets to form a first training set, where k < M and M and k are both positive integers;

[0097] The preprocessing unit is used to preprocess the first training set by using at least one processing method of color conversion processing, flip processing and clipping processing to obtain the preprocessed target training set.

[0098] Preferably, the preprocessing unit is used to perform trimming processing on the first training set to obtain data of multiple sub-images corresponding to the first image data;

[0099] The data of each sub-image constitutes a target training set.

[0100] Preferably, the preprocessing unit is used to perform trimming processing on the first training set to obtain data of multiple sub-images corresponding to the first image data;

[0101] The first prediction vector acquisition module is further used to use the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image;

[0102] The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process;

[0103] The first feature vector acquisition module is further used to acquire the first feature vector corresponding to the residual module obtained by the residual network in the process of extracting attribute features at different levels of data of each sub-image;

[0104] The second feature vector acquisition module is also used to use the pooling method to perform pooling processing on the first feature vectors corresponding to the data of each sub-image at different levels, so as to obtain the second feature vectors corresponding to the data of each sub-image at different levels.

[0105] Preferably, after the residual network is used to perform attribute feature extraction processing at different levels on the data of each sub-image, the first prediction vector acquisition module includes a first sub-prediction vector acquisition unit and a first average aggregation processing unit;

[0106] The first sub-prediction vector acquisition unit is used to acquire the first sub-prediction vector corresponding to the data of each sub-image at different levels;

[0107] The first average aggregation processing unit is used to perform average aggregation processing on the first sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the first attribute prediction vector corresponding to the data of each sub-image;

[0108] The intermediate prediction vector acquisition module includes a second sub-prediction vector acquisition unit, a third sub-prediction vector acquisition unit, a second average aggregation processing unit and a third average aggregation processing unit;

[0109] The second sub-prediction vector acquisition unit is used to use the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels;

[0110] The second average aggregation processing unit is used to perform average aggregation processing on the second sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the second attribute prediction vector corresponding to the data of each sub-image;

[0111] The third sub-prediction vector acquisition unit is used to use the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the third sub-prediction vector corresponding to the data of each sub-image at different levels;

[0112] The third average aggregation processing unit is used to perform average aggregation processing on the third sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the third attribute prediction vector corresponding to the data of each sub-image;

[0113] The prediction model building module is also used to perform splicing processing on the intermediate prediction models corresponding to the data of each sub-image to establish the target attribute prediction model for predicting product attributes.

[0114] Preferably, the prediction model building module is also used to splice the intermediate prediction models corresponding to the data of each sub-image when the intermediate prediction models corresponding to the data of each sub-image have equal weights, so as to establish the target attribute prediction model for predicting product attributes.

[0115] Preferably, after the residual network is used to perform attribute feature extraction processing at different levels on the data of each sub-image, the first prediction vector acquisition module includes a sub-prediction vector acquisition unit and a first regression processing unit;

[0116] The first sub-prediction vector acquisition unit is used to acquire the first sub-prediction vector corresponding to the data of each sub-image at different levels;

[0117] The first regression processing unit is used to use a regression algorithm to perform regression processing on the first sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the first attribute prediction vector corresponding to the first image data;

[0118] The intermediate prediction vector acquisition module includes a second sub-prediction vector acquisition unit, a third sub-prediction vector acquisition unit, a first regression processing unit and a second regression processing unit;

[0119] The second sub-prediction vector acquisition unit is used to use the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels;

[0120] The first regression processing unit is used to use the regression algorithm to perform regression processing on the second sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the second attribute prediction vector corresponding to the first image data;

[0121] The third sub-prediction vector acquisition unit is used to use the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the third sub-prediction vector corresponding to the data of each sub-image at different levels;

[0122] The second regression processing unit is used to use the regression algorithm to perform regression processing on the third sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the third attribute prediction vector corresponding to the first image data.

[0123] Preferably, the first prediction vector acquisition module includes a convolution processing unit, a pooling processing unit and a third regression processing unit;

[0124] The convolution processing unit is used to perform convolution processing on the target training set using a residual network to obtain N residual modules, where N is a positive integer;

[0125] The pooling processing unit is used to perform pooling processing on the Nth residual module using the pooling method to obtain an intermediate feature vector;

[0126] The dimension of the intermediate feature vector is higher than the dimension of the first feature vector corresponding to the residual module;

[0127] The third regression processing unit is used to perform regression processing on the intermediate feature vector using a regression algorithm to obtain the first attribute prediction vector.

[0128] Preferably, when the first prediction vector acquisition module adopts a residual network to include N residual modules in the attribute feature extraction process, the first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process.

[0129] Preferably, the residual network includes a deep residual network or Inception; and / or,

[0130] The linear classifier comprises a support vector machine; and / or,

[0131] The nonlinear classifier includes a regression classification tree or a random forest algorithm; and / or,

[0132] The pooling method includes global average pooling; and / or,

[0133] The regression algorithm includes a multi-class logistic regression algorithm.

[0134] Preferably, the target attribute prediction module includes a selection unit and a determination unit;

[0135] The selection unit is used to input the second image data into the target attribute prediction model, and select the maximum vector value in the prediction vector corresponding to the target attribute prediction model;

[0136] The determining unit is used to determine that the attribute corresponding to the maximum vector value is the target attribute of the target product.

[0137] Preferably, the prediction system further includes a verification set acquisition module;

[0138] The verification set acquisition module is used to select Mk pieces of the sub-datasets to form a verification set;

[0139] The prediction system also includes an actual attribute acquisition module, a verification module and a judgment module;

[0140] The actual attribute acquisition module is used to acquire the actual attribute of the first product being trained;

[0141] The verification module is used to input the sub-dataset in the verification set into the target attribute prediction model to obtain the prediction result of the first product;

[0142] The judgment module is used to judge whether the prediction result is consistent with the actual attribute of the trained product. If so, it is determined that the target attribute prediction model is feasible; otherwise, the target attribute prediction model is retrained until the prediction result is consistent with the actual attribute of the first product being trained.

[0143] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for predicting product attributes when executing the computer program.

[0144] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and the computer program, when executed by a processor, implements the steps of the above-mentioned product attribute prediction method.

[0145] The positive and progressive effects of the present invention are:

[0146] In the present invention, on the one hand, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as a basic model, and a basic model group is used to replace the basic model used in the usual solution to better capture the multi-layer semantic knowledge contained in the image. Each basic model works at a semantic level, and the entire basic model group works at a multi-layer semantic level, thereby learning more image data and having stronger prediction capabilities; on the other hand, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually fuse heterogeneous models, thereby improving nonlinear capabilities, reducing feature dimensions, and improving overall learning efficiency, thereby overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0147] Figure 1 This is a flow chart of the method for predicting product attributes of Example 1 of the present invention.

[0148] Figure 2 This is a flow chart of a method for predicting product attributes according to Embodiment 2 of the present invention.

[0149] Figure 3 This is a flow chart of a method for predicting product attributes according to Embodiment 3 of the present invention.

[0150] Figure 4 This is a flow chart of a method for predicting product attributes according to Embodiment 4 of the present invention.

[0151] Figure 5 This is a flow chart of the method for predicting product attributes of Example 5 of the present invention.

[0152] Figure 6 This is a module diagram of a system for predicting product attributes according to Example 6 of the present invention.

[0153] Figure 7 This is a module diagram of a system for predicting product attributes according to Embodiment 7 of the present invention.

[0154] Figure 8 This is a module diagram of a system for predicting product attributes according to Embodiment 8 of the present invention.

[0155] Fig. 9 This is a module diagram of a system for predicting product attributes according to Embodiment 9 of the present invention.

[0156] Fig.10 This is a module diagram of a system for predicting product attributes according to Embodiment 10 of the present invention.

[0157] Fig.11This is a schematic diagram of the structure of an electronic device for implementing the method for predicting product attributes in Example 11 of the present invention. DETAILED DESCRIPTION

[0158] The present invention is further described below by way of examples, but the present invention is not limited to the scope of the examples.

[0159] Example 1

[0160] like Figure 1 As shown, the method for predicting product attributes of this embodiment includes:

[0161] S101, obtaining first image data of a first product to be trained;

[0162] S102, acquiring a target training set according to the first image data;

[0163] S103, using a residual network to perform attribute feature extraction processing on the target training set, and obtaining a first attribute prediction vector output by the residual network;

[0164] Specifically, a residual network is used to perform convolution processing on the target training set to obtain N residual modules, where N is a positive integer;

[0165] The pooling method is used to perform pooling processing on the Nth residual module to obtain the intermediate feature vector;

[0166] The regression algorithm is used to perform regression processing on the intermediate feature vector to obtain the first attribute prediction vector.

[0167] Preferably, when the residual network includes N residual modules in the attribute feature extraction process, the first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process.

[0168] S104, obtaining a first feature vector corresponding to the residual module obtained by the residual network in the attribute feature extraction process;

[0169] Among them, the dimension of the intermediate feature vector is higher than the dimension of the first feature vector corresponding to the residual module;

[0170] S105, performing pooling processing on the first feature vector using a pooling method to obtain a second feature vector;

[0171] S106, using a linear classifier and a nonlinear classifier to classify the second feature vector to obtain an intermediate attribute prediction vector;

[0172] S107, concatenating the first attribute prediction vector and the intermediate attribute prediction vector to establish a target attribute prediction model for predicting product attributes;

[0173] S108, obtaining second image data of the target product to be predicted;

[0174] S109: Input the second image data into a target attribute prediction model to predict target attributes of the target product.

[0175] In this embodiment, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as the basic model, which can learn more image data and have stronger prediction capabilities. At the same time, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually merge, thereby improving nonlinear capabilities and reducing feature dimensions, thereby improving overall learning efficiency, overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities.

[0176] Example 2

[0177] like Figure 2 As shown, the prediction method of product attributes in this embodiment is a further improvement on Embodiment 1, specifically:

[0178] When different levels of information of multiple sub-images corresponding to the first image data of the product are used as training objects, step S102 includes:

[0179] S1021, dividing the first image data into M sub-datasets, and selecting k sub-datasets to form a first training set, where k < M and M and k are both positive integers;

[0180] S1022, preprocessing the first training set by using at least one of color conversion, flipping, and cropping to obtain a preprocessed target training set;

[0181] When the first training set is trimmed, data of a plurality of sub-images corresponding to the first image data are obtained.

[0182] When different levels of information of multiple sub-images corresponding to the first image data of the product are used as training objects, step S103 includes:

[0183] S1031, using a residual network to perform attribute feature extraction processing at different levels on the data of each sub-image, and obtaining a first attribute prediction vector output by the residual network;

[0184] Among them, the deeper the layer, the more residual modules the residual network corresponds to in the attribute feature extraction process;

[0185] Step S104 includes:

[0186] S1041, obtaining a first feature vector corresponding to a residual module obtained by the residual network in a process of extracting attribute features at different levels of data of each sub-image;

[0187] Step S105 includes:

[0188] S1051 , using a pooling method to perform pooling processing on the first eigenvectors corresponding to the data of each sub-image at different levels, to obtain the second eigenvectors corresponding to the data of each sub-image at different levels.

[0189] After step S105 and before step S106, the following steps are also included:

[0190] S1060: Use principal component analysis technology to perform dimensionality reduction processing on the second eigenvector to obtain the second eigenvector after dimensionality reduction processing.

[0191] Step S106 includes:

[0192] S1061, classify the second feature vector using a linear classifier to obtain a second attribute prediction vector;

[0193] S1062, classify the second feature vector using a nonlinear classifier to obtain a third attribute prediction vector;

[0194] Step S107 includes:

[0195] S1071. Concatenate the first attribute prediction vector, the second attribute prediction vector, and the third attribute prediction vector to establish a target attribute prediction model for predicting product attributes.

[0196] Step S109 includes:

[0197] S1091, inputting the second image data into the target attribute prediction model, and selecting the maximum vector value in the prediction vector corresponding to the target attribute prediction model;

[0198] S1092. Determine the attribute corresponding to the maximum vector value as the target attribute of the target product.

[0199] Specifically, the input image to be predicted is extracted into 27 sampling sub-images according to the fixed sampling method in the preprocessing method, and other preprocessing is performed. Then, they are used as inputs of the integrated model (including models of multiple groups of model groups). Each sub-image can eventually obtain a prediction attribute vector. Then, the 27 prediction attribute vectors are voted with equal weights to obtain the attribute with the highest proportion in the prediction attribute vector as the final prediction attribute of the input image.

[0200] Among them, residual networks include deep residual networks, Inception, etc.

[0201] Linear classifiers include support vector machines, etc.

[0202] Nonlinear classifiers include regression classification trees, random forest algorithms, etc.

[0203] Pooling methods include global average pooling, etc.

[0204] Regression algorithms include multi-class logistic regression algorithms, etc.

[0205] In addition, when the first image data of the product is directly used as the training object, step S102 includes:

[0206] The first image data is divided into M sub-datasets, and k sub-datasets are selected to form a target training set, where k < M and M and k are both positive integers.

[0207] When the data of a plurality of sub-images corresponding to the first image data of the product are used as training objects, step S102 includes:

[0208] Divide the first image data into M sub-datasets, and select k sub-datasets to form a first training set, where k < M and M and k are both positive integers;

[0209] The first training set is preprocessed by using at least one of color conversion processing, flip processing and cropping processing to obtain a preprocessed target training set.

[0210] Specifically, when the first training set is trimmed, data of multiple sub-images corresponding to the first image data are obtained; wherein the data of each sub-image constitutes a target training set.

[0211] In this embodiment, the case where different levels of information of multiple sub-images corresponding to the first image data of the product are used as training objects is mainly explained. By analogy, it can be known that the processing procedures of the two cases of directly using the first image data of the product as the training object and using the data of multiple sub-images corresponding to the first image data of the product as the training object are two different cases, so they will not be repeated here.

[0212] In this embodiment, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as the basic model, so as to better capture the multi-layer semantic knowledge contained in the image, learn more picture data, and have stronger prediction ability; at the same time, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually merge, thereby improving nonlinear capabilities, reducing feature dimensions, and improving overall learning efficiency. It overcomes the serious overfitting and high prediction errors of existing prediction methods, improves the prediction accuracy of product attributes, and the entire method has good robustness and practical application capabilities.

[0213] Example 3

[0214] like Figure 3 As shown, the prediction method of product attributes in this embodiment is a further improvement of Embodiment 2, specifically:

[0215] Step S1031 includes:

[0216] S10311, using a residual network to perform attribute feature extraction processing at different levels on the data of each sub-image;

[0217] S10312, obtaining the first sub-prediction vector corresponding to the data of each sub-image at different levels;

[0218] S10313, respectively perform average aggregation processing on the first sub-prediction vectors corresponding to the data of the same sub-image at different levels to obtain a first attribute prediction vector corresponding to the data of each sub-image;

[0219] Step S1061 includes:

[0220] A linear classifier is used to classify the second feature vectors corresponding to the data of each sub-image at different levels, and the second sub-prediction vectors corresponding to the data of each sub-image at different levels are obtained;

[0221] Performing average aggregation processing on the second sub-prediction vectors corresponding to the data of the same sub-image at different levels respectively, to obtain the second attribute prediction vector corresponding to the data of each sub-image;

[0222] Step S1062 includes:

[0223] A nonlinear classifier is used to classify the second feature vectors corresponding to the data of each sub-image at different levels, and a third sub-prediction vector corresponding to the data of each sub-image at different levels is obtained;

[0224] Performing average aggregation processing on the third sub-prediction vectors corresponding to the data of the same sub-image at different levels, respectively, to obtain the third attribute prediction vector corresponding to the data of each sub-image;

[0225] Step S1071 includes:

[0226] The first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector are concatenated to obtain an intermediate prediction model corresponding to the data of each sub-image;

[0227] The intermediate prediction models corresponding to the data of each sub-image are spliced ​​to establish a target attribute prediction model for predicting product attributes.

[0228] Wherein, when the intermediate prediction models corresponding to the data of each sub-image have equal weights, the intermediate prediction models corresponding to the data of each sub-image are spliced ​​to establish a target attribute prediction model for predicting product attributes.

[0229] The following is a specific example:

[0230] 1) Divide the image data

[0231] The image data is divided into M parts. Each time the model is trained, (M-1) parts are used as the training set, and the remaining 1 part is used as the validation set.

[0232] 2) Preprocess the image data

[0233] Perform data enhancement preprocessing on the image, such as changing color, flipping left and right, cropping, etc.;

[0234] As for image cropping, according to actual observation experience, the distribution of the center position of the products in the images of most products on the Internet platform has a strong regularity, such as being concentrated near the center of the image or the vertical center line, and the size of the products does not vary greatly; therefore, the following random cropping scheme can be used to crop the image: scale the width and height of the image to a set size (such as 224 pixels), and then construct an 8×8 regular grid on the image (such as Figure 4 As shown in the figure, three square areas with central side lengths of 56, 112, and 168 (represented by a, b, and c, respectively) are taken as random cropping centers, and the cropping sizes are 168, 112, and 56, respectively. The horizontal positions are sampled according to the normal distribution (that is, the image center is the mean and the variance is σ), and the vertical positions are sampled according to the uniform distribution. This sampling method can effectively integrate the "attention" prior distribution. In addition, for each type of cropping size, the four corners of the square and the 7 points intersecting the vertical center axis of the image are retained as fixed sampling points, so that a total of 27 fixed cropping methods are used for both training and testing.

[0235] 3) Establish basic model and basic model group

[0236] The data of each sub-image is processed at different levels including:

[0237] (1) Use a deep residual network (such as Resnetv2) to perform convolution processing on the data of sub-images at different levels, obtain N residual modules, and obtain the first attribute prediction vector output by the residual network, where N is a positive integer;

[0238] Perform GlobalAvgPool (global average pooling) on ​​the Nth residual module to obtain the intermediate feature vector;

[0239] The Softmax regression algorithm is used to regress the intermediate feature vector to obtain the first attribute prediction vector;

[0240] (2) The first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process, and the first feature vector is subjected to global average pooling processing to obtain a second feature vector;

[0241] PCA (principal component analysis) is used to perform dimensionality reduction processing on the second eigenvector to obtain the second eigenvector after dimensionality reduction processing;

[0242] Among them, some features were retained through principal component analysis technology, and the variance was retained at 95%.

[0243] A linear model SVM (support vector machine) is used to classify the second feature vector after dimension reduction to obtain a second attribute prediction vector;

[0244] The nonlinear model CART (regression classification tree) is used to classify the second eigenvector after dimensionality reduction to obtain the third attribute prediction vector.

[0245] The formula corresponding to the first attribute prediction vector is as follows:

[0246]

[0247] The product has m attributes, w i represents the first training parameter, w j represents the second training parameter, x represents different attributes, P i Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute, i, j = 0, 1, ..., m and are all integers;

[0248] The formula corresponding to the second attribute prediction vector is as follows:

[0249] y ij =w ij x+b ij ,

[0250] w ij represents the third training parameter, b ij represents the fourth training parameter, y ij Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute;

[0251] The formula corresponding to the third attribute prediction vector is as follows:

[0252]

[0253] N iRepresents the number of samples of class i in the leaf node, N j represents the number of samples of class j in the leaf node, P i Represents the prediction vector corresponding to class i (i.e., the probability corresponding to class i);

[0254] Among them, the first attribute prediction vector obtained by the Softmax layer of the deep residual network Resnetv2 is used as the Softmax main classifier, the SVM auxiliary classifier of the second attribute prediction vector obtained by SVM and the third attribute prediction vector obtained by CART are used as CART auxiliary classifiers, and then the main classifier and the auxiliary classifier are combined to obtain the final classification result, thereby effectively improving the classification performance based on a single model.

[0255] (3) For different levels of data of each sub-image, corresponding basic models are constructed respectively according to the above steps (1) and (2), and a basic model group corresponding to the data of each sub-image is constructed based on the basic model.

[0256] For example, it is assumed that the six levels of the data of each sub-image are trained separately, then the data of each sub-image corresponds to six basic models, and the six basic models together constitute a basic model group, that is, the data of each sub-image corresponds to a basic model group.

[0257] The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process, and the more abstract the information it contains, thereby capturing the multi-layer semantic knowledge contained in the image and learning more image data. Generally, 3-8 residual modules are used.

[0258] Each basic model contains three classifiers, namely, Softmax main classifier, SVM auxiliary classifier and CART auxiliary classifier, so each basic model group includes 18 classifiers, that is, the data of each sub-image corresponds to 18 classifiers.

[0259] For a basic model group, each basic model outputs the first sub-prediction vector corresponding to the Softmax main classifier, the second sub-prediction vector corresponding to the SVM auxiliary classifier, and the third sub-prediction vector corresponding to the CART auxiliary classifier; then the first sub-prediction vectors of different basic models are averaged and aggregated to obtain the first attribute prediction vector corresponding to the sub-image data; the second sub-prediction vectors of different basic models are averaged and aggregated to obtain the second attribute prediction vector corresponding to the sub-image data; the third sub-prediction vectors of different basic models are averaged and aggregated to obtain the third attribute prediction vector corresponding to the sub-image data, so as to obtain three attribute prediction vectors corresponding to a basic model group, and the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector are spliced ​​to obtain an intermediate prediction model corresponding to a basic model group; finally, the intermediate prediction models corresponding to each basic model group are spliced ​​to establish a target attribute prediction model for predicting product attributes, thereby predicting product attributes.

[0260] In this embodiment, on the one hand, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as a basic model, and a basic model group is used to replace the basic model used in the usual solution to better capture the multi-layer semantic knowledge contained in the image. Each basic model works at a semantic level, and the entire basic model group works at a multi-layer semantic level, thereby learning more image data and having stronger prediction capabilities; on the other hand, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually merge, thereby improving nonlinear capabilities, reducing feature dimensions, and improving overall learning efficiency, thereby overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities.

[0261] Example 4

[0262] like Figure 4 As shown, the prediction method of product attributes in this embodiment is a further improvement of Embodiment 2, specifically:

[0263] Step S1031 includes:

[0264] S10311, using a residual network to perform attribute feature extraction processing at different levels on the data of each sub-image;

[0265] S10312, obtaining the first sub-prediction vector corresponding to the data of each sub-image at different levels;

[0266] S10314, using a regression algorithm to perform regression processing on first sub-prediction vectors corresponding to data of different sub-images at different levels, to obtain a first attribute prediction vector corresponding to the first image data;

[0267] Step S1061 includes:

[0268] A linear classifier is used to classify the second feature vectors corresponding to the data of each sub-image at different levels, and the second sub-prediction vectors corresponding to the data of each sub-image at different levels are obtained;

[0269] Using a regression algorithm to perform regression processing on the second sub-prediction vectors corresponding to the data of different sub-images at different levels, to obtain a second attribute prediction vector corresponding to the first image data;

[0270] Step S1062 includes:

[0271] A nonlinear classifier is used to classify the second feature vectors corresponding to the data of each sub-image at different levels, and a third sub-prediction vector corresponding to the data of each sub-image at different levels is obtained;

[0272] A regression algorithm is used to perform regression processing on the third sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the third attribute prediction vector corresponding to the first image data.

[0273] The following is a specific example:

[0274] 1) Divide the image data

[0275] The image data is divided into M parts. Each time the model is trained, (M-1) parts are used as the training set, and the remaining 1 part is used as the validation set.

[0276] 2) Preprocess the image data

[0277] Perform data enhancement preprocessing on the image, such as changing color, flipping left and right, cropping, etc.;

[0278] As for image cropping, according to actual observation experience, the distribution of the center position of the products in the images of most products on the Internet platform has a strong regularity, such as being concentrated near the center of the image or the vertical center line, and the size of the products does not vary greatly; therefore, the following random cropping scheme can be used to crop the image: scale the width and height of the image to a set size (such as 224 pixels), and then construct an 8×8 regular grid on the image (such as Figure 4As shown in the figure, three square areas with central side lengths of 56, 112, and 168 (represented by a, b, and c, respectively) are taken as random cropping centers, and the cropping sizes are 168, 112, and 56, respectively. The horizontal positions are sampled according to the normal distribution (that is, the image center is the mean and the variance is σ), and the vertical positions are sampled according to the uniform distribution. This sampling method can effectively integrate the "attention" prior distribution. In addition, for each type of cropping size, the four corners of the square and the 7 points intersecting the vertical center axis of the image are retained as fixed sampling points, so that a total of 27 fixed cropping methods are used for both training and testing.

[0279] 3) Establish basic model and basic model group

[0280] The data of each sub-image is processed at different levels including:

[0281] (1) Use a deep residual network (such as Resnetv2) to perform convolution processing on the data of sub-images at different levels, obtain N residual modules, and obtain the first attribute prediction vector output by the residual network, where N is a positive integer;

[0282] Perform GlobalAvgPool (global average pooling) on ​​the Nth residual module to obtain the intermediate feature vector;

[0283] The Softmax regression algorithm is used to regress the intermediate feature vector to obtain the first attribute prediction vector;

[0284] (2) The first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process, and the first feature vector is subjected to global average pooling processing to obtain a second feature vector;

[0285] PCA (principal component analysis) is used to perform dimensionality reduction processing on the second eigenvector to obtain the second eigenvector after dimensionality reduction processing;

[0286] Among them, some features were retained through principal component analysis technology, and the variance was retained at 95%.

[0287] A linear model SVM (support vector machine) is used to classify the second feature vector after dimension reduction to obtain a second attribute prediction vector;

[0288] The nonlinear model CART (regression classification tree) is used to classify the second eigenvector after dimensionality reduction to obtain the third attribute prediction vector.

[0289] The formula corresponding to the first attribute prediction vector is as follows:

[0290]

[0291] The product has m attributes, w i represents the first training parameter, w j represents the second training parameter, x represents different attributes, P i Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute, i, j = 0, 1, ..., m and are all integers;

[0292] The formula corresponding to the second attribute prediction vector is as follows:

[0293] y ij =w ij x+b ij ,

[0294] w ij represents the third training parameter, b ij represents the fourth training parameter, y ij Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute;

[0295] The formula corresponding to the third attribute prediction vector is as follows:

[0296]

[0297] N i Indicates that N j Indicates that P i Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute;

[0298] Among them, the first attribute prediction vector obtained by the Softmax layer of the deep residual network Resnetv2 is used as the Softmax main classifier, the SVM auxiliary classifier of the second attribute prediction vector obtained by SVM and the third attribute prediction vector obtained by CART are used as CART auxiliary classifiers, and then the main classifier and the auxiliary classifier are combined to obtain the final classification result, thereby effectively improving the classification performance based on a single model.

[0299] (3) For different levels of data of each sub-image, corresponding basic models are constructed respectively according to the above steps (1) and (2), and a basic model group corresponding to the data of each sub-image is constructed based on the basic model.

[0300] For example, if the six levels of data for each sub-image are trained separately, then the data for each sub-image corresponds to six basic models, which together form a basic model group, that is, the data for each sub-image corresponds to a basic model group. The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process, and the more abstract the information contained, thereby capturing the multi-layer semantic knowledge contained in the image and learning more image data. Generally, 3-8 residual modules are used.

[0301] Each basic model contains three classifiers, namely, Softmax main classifier, SVM auxiliary classifier and CART auxiliary classifier, so each basic model group includes 18 classifiers, that is, the data of each sub-image corresponds to 18 classifiers.

[0302] For a basic model group, each basic model outputs the first sub-prediction vector corresponding to the Softmax main classifier, the second sub-prediction vector corresponding to the SVM auxiliary classifier, and the third sub-prediction vector corresponding to the CART auxiliary classifier; then the first sub-prediction vectors in different basic model groups are regressed to obtain the first attribute prediction vector corresponding to the image data; the second sub-prediction vectors in different basic model groups are regressed to obtain the second attribute prediction vector corresponding to the image data; the third sub-prediction vectors in different basic model groups are regressed to obtain the third attribute prediction vector corresponding to the sub-image data, so as to obtain the three attribute prediction vectors corresponding to the image data, and the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector are spliced ​​to obtain, and a target attribute prediction model for predicting product attributes is established to further predict product attributes.

[0303] In this embodiment, on the one hand, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as a basic model, and a basic model group is used to replace the basic model used in the usual solution to better capture the multi-layer semantic knowledge contained in the image. Each basic model works at a semantic level, and the entire basic model group works at a multi-layer semantic level, thereby learning more image data and having stronger prediction capabilities; on the other hand, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually fuse heterogeneous models, thereby improving nonlinear capabilities, reducing feature dimensions, and improving overall learning efficiency, thereby overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities.

[0304] Example 5

[0305] like Figure 5As shown, the prediction method of product attributes in this embodiment is a further improvement on Embodiment 2, specifically:

[0306] Step S102 also includes:

[0307] The Mk molecular data set was selected to form the validation set;

[0308] The data in the validation set are preprocessed, and a basic model group is trained on each training set, with a total of m basic model groups trained. Each basic model group contains 6 basic models, which have a similar structure, except that the number of residuals of the resnetv2 model is different, generally 3 to 8. The feature map outputs of the last two residual models are used to extract features as training samples for SVM and CART, and the model hyperparameters are selected through the validation set.

[0309] The validation set is enhanced using data preprocessing methods, and then m basic model groups are input respectively. Each basic model in each basic model group outputs three types of prediction vectors: Softmax, SVM, and CART. The outputs of all homogeneous models in the basic model group are averaged and aggregated, so that three prediction vectors are obtained as the feature representation of this model group.

[0310] Different model groups obtain different feature representations. Assuming that there are c categories of clothing attributes, the output feature dimension of each basic model group is [c+c(c-1) / 2+c], and the total feature representation dimension obtained by the basic model layer is [mc(c+3) / 2]; each intermediate classifier in the middle layer uses part of the features as input, and the input dimensions of each intermediate classifier are mc, [mc(c-1) / 2], mc, respectively, and the corresponding attributes are the same as the original input image.

[0311] Specifically, the output dimension of CART in the basic model is c, and the feature value is the distribution of c-type samples in the leaf node into which the sample falls.

[0312] For example, if the number of samples of class c contained in the leaf node is m1, m2, ..., m c , then the output prediction vector is (m1 / m, m2 / m, …, m c / m), where The output dimension of the SVM in the basic model is [c(c-1) / 2]. Since SVM is a binary classifier, in order to solve the c-class problem, it is necessary to train [c(c-1) / 2] classifiers, that is, c i c j A classifier is trained for every two types of samples, and the prediction value of each classifier is used as the feature value of that dimension; the softmax prediction value is used as the c-dimensional feature vector.

[0313] Therefore, the feature outputs of the homogeneous models in each basic model group are first averaged and aggregated, that is, the feature dimensions obtained after the average aggregation of the homogeneous model SVM, softmax, and CART outputs in the k basic models are c, c(c-1) / 2, c;

[0314] The output features of the homogeneous models in different basic model groups are concatenated. Assuming there are m basic model groups, the feature input dimensions of the intermediate classifier corresponding to each homogeneous model are mc, [mc(c-1) / 2], mc, and the output dimension is c. The input is a 3-category heterogeneous model with a dimension of 3c, and the output is a c-dimensional prediction probability vector;

[0315] All of the above classifiers use Softmax as the weight learner of the combined model, avoiding a simple voting mechanism and enhancing the predictive ability of the overall model.

[0316] After step S109, the following steps are also included:

[0317] S1010, obtaining actual attributes of the first product being trained;

[0318] S1011, inputting the sub-data set in the validation set into the target attribute prediction model to obtain the prediction result of the first product;

[0319] S1012. Determine whether the prediction result is consistent with the actual attribute of the trained product. If so, determine that the target attribute prediction model is feasible; otherwise, retrain the target attribute prediction model until the prediction result is consistent with the actual attribute of the first trained product.

[0320] In this embodiment, the accuracy of the target attribute prediction model is verified through a validation set to ensure the accuracy of the target attribute prediction model in predicting product attributes, so that the entire method has good robustness and practical application capabilities.

[0321] Example 6

[0322] like Figure 6 As shown, the product attribute prediction system of this embodiment includes a first image data acquisition module 1, a target training set acquisition module 2, a first prediction vector acquisition module 3, a first feature vector acquisition module 4, a second feature vector acquisition module 5, an intermediate prediction vector acquisition module 6, a prediction model establishment module 7, a second image data acquisition module 8 and a target attribute prediction module 9;

[0323] The first image data acquisition module 1 is used to acquire first image data of a first product to be trained;

[0324] The target training set acquisition module 2 is used to acquire the target training set according to the first image data;

[0325] The first prediction vector acquisition module 3 is used to use the residual network to perform attribute feature extraction processing on the target training set to obtain the first attribute prediction vector output by the residual network;

[0326] Specifically, the first prediction vector acquisition module includes a convolution processing unit, a pooling processing unit and a third regression processing unit;

[0327] The convolution processing unit is used to perform convolution processing on the target training set using a residual network to obtain N residual modules, where N is a positive integer;

[0328] The pooling processing unit is used to perform pooling processing on the Nth residual module using a pooling method to obtain an intermediate feature vector;

[0329] The third regression processing unit is used to perform regression processing on the intermediate feature vector using a regression algorithm to obtain a first attribute prediction vector.

[0330] Specifically, when the first prediction vector acquisition module uses a residual network to include N residual modules in the attribute feature extraction process, the first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process.

[0331] The first feature vector acquisition module 4 is used to obtain the first feature vector corresponding to the residual module obtained by the residual network during the attribute feature extraction process;

[0332] Among them, the dimension of the intermediate feature vector is higher than the dimension of the first feature vector corresponding to the residual module;

[0333] The second feature vector acquisition module 5 is used to perform pooling processing on the first feature vector by using a pooling method to obtain a second feature vector;

[0334] The intermediate prediction vector acquisition module 6 is used to classify the second feature vector using a linear classifier and a nonlinear classifier to obtain an intermediate attribute prediction vector;

[0335] The prediction model building module 7 is used to perform splicing processing on the first attribute prediction vector and the intermediate attribute prediction vector to establish a target attribute prediction model for predicting product attributes;

[0336] The second image data acquisition module 8 is used to acquire the second image data of the target product to be predicted;

[0337] The target attribute prediction module 9 is used to input the second image data into the target attribute prediction model to predict the target attribute of the target product.

[0338] In this embodiment, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as the basic model, which can learn more image data and have stronger prediction capabilities. At the same time, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually merge, thereby improving nonlinear capabilities and reducing feature dimensions, thereby improving overall learning efficiency, overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities.

[0339] Example 7

[0340] like Figure 7 As shown, the product attribute prediction system of this embodiment is a further improvement of the embodiment 6, specifically:

[0341] When different levels of information of multiple sub-images corresponding to the first image data of the product are used as training objects, the target training set acquisition module 2 includes a training set acquisition unit 21 and a preprocessing unit 22 .

[0342] The training set acquisition unit 21 is used to divide the first image data into M sub-data sets, and select k sub-data sets to form a first training set, where k < M and M and k are both positive integers;

[0343] The preprocessing unit 22 is used to preprocess the first training set by using at least one of color conversion processing, flip processing and cropping processing to obtain a preprocessed target training set.

[0344] Wherein, when the preprocessing unit 22 is used to perform a trimming process on the first training set, data of a plurality of sub-images corresponding to the first image data are obtained;

[0345] The first prediction vector acquisition module 3 is also used to use a residual network to perform attribute feature extraction processing at different levels on the data of each sub-image;

[0346] Among them, the deeper the layer, the more residual modules the residual network corresponds to in the attribute feature extraction process;

[0347] The first feature vector acquisition module 4 is also used to obtain the first feature vector corresponding to the residual module obtained by the residual network in the process of extracting attribute features at different levels of data of each sub-image;

[0348] The second feature vector acquisition module 5 is further used to perform pooling processing on the first feature vectors corresponding to the data of each sub-image at different levels by using a pooling method, so as to acquire the second feature vectors corresponding to the data of each sub-image at different levels.

[0349] The prediction system also includes a principal component analysis module 10;

[0350] The principal component analysis module 10 is used to perform dimensionality reduction processing on the second eigenvector by using the principal component analysis technology to obtain the second eigenvector after dimensionality reduction processing.

[0351] The intermediate prediction vector acquisition module 6 is used to classify the second feature vector using a linear classifier to obtain a second attribute prediction vector;

[0352] The intermediate prediction vector acquisition module 6 is further used to classify the second feature vector using a nonlinear classifier to obtain a third attribute prediction vector;

[0353] The prediction model building module 7 is used to perform splicing processing on the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector to establish a target attribute prediction model for predicting product attributes.

[0354] Wherein, the target attribute prediction module 9 includes a selection unit 91 and a determination unit 92;

[0355] The selection unit 91 is used to input the second image data into the target attribute prediction model, and select the maximum vector value in the prediction vector corresponding to the target attribute prediction model;

[0356] The determination unit 92 is used to determine the attribute corresponding to the maximum vector value as the target attribute of the target product.

[0357] Specifically, the input image to be predicted is extracted into 27 sampling sub-images according to the fixed sampling method in the preprocessing method, and other preprocessing is performed. Then, they are used as inputs of the integrated model (including models of multiple groups of model groups). Each sub-image can eventually obtain a prediction attribute vector. Then, the 27 prediction attribute vectors are voted with equal weights to obtain the attribute with the highest proportion in the prediction attribute vector as the final prediction attribute of the input image.

[0358] Among them, residual networks include deep residual networks, Inception, etc.

[0359] Linear classifiers include support vector machines, etc.

[0360] Nonlinear classifiers include regression classification trees, random forest algorithms, etc.

[0361] Pooling methods include global average pooling, etc.

[0362] Regression algorithms include multi-class logistic regression algorithms, etc.

[0363] In addition, when the first image data of the product is directly used as the training object, the target training set acquisition module is used to divide the first image data into M sub-datasets, and select k sub-datasets to form the target training set, k<M and M and k are both positive integers.

[0364] When the data of multiple sub-images corresponding to the first image data of the product are used as training objects, the target training set acquisition module 2 includes a training set acquisition unit and a preprocessing unit.

[0365] The training set acquisition unit is used to divide the first image data into M sub-data sets, and select k sub-data sets to form a first training set, where k < M and M and k are both positive integers;

[0366] The preprocessing unit is used to preprocess the first training set by using at least one of color conversion processing, flip processing and clipping processing to obtain a preprocessed target training set.

[0367] When the preprocessing unit is used to perform a trimming process on the first training set, data of multiple sub-images corresponding to the first image data are obtained; and the data of each sub-image constitutes a target training set.

[0368] In this embodiment, the case where different levels of information of multiple sub-images corresponding to the first image data of the product are used as training objects is mainly explained. By analogy, it can be known that the processing procedures of the two cases of directly using the first image data of the product as the training object and using the data of multiple sub-images corresponding to the first image data of the product as the training object are two different cases, so they will not be repeated here.

[0369] In this embodiment, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as the basic model, which can learn more image data and have stronger prediction capabilities. At the same time, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually merge, thereby improving nonlinear capabilities and reducing feature dimensions, thereby improving overall learning efficiency, overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities.

[0370] Example 8

[0371] like Figure 8 As shown, the product attribute prediction system of this embodiment is a further improvement of Embodiment 7, specifically:

[0372] After using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image, the first prediction vector acquisition module 3 includes a first sub-prediction vector acquisition unit 31 and a first average aggregation processing unit 32;

[0373] The first sub-prediction vector acquisition unit 31 is used to acquire the first sub-prediction vector corresponding to the data of each sub-image at different levels;

[0374] The first average aggregation processing unit 32 is used to perform average aggregation processing on the first sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the first attribute prediction vector corresponding to the data of each sub-image;

[0375] The intermediate prediction vector acquisition module 6 includes a second sub-prediction vector acquisition unit 61, a third sub-prediction vector acquisition unit 62, a second average aggregation processing unit 63 and a third average aggregation processing unit 64;

[0376] The second sub-prediction vector acquisition unit 61 is used to classify the second feature vectors corresponding to the data of each sub-image at different levels by using a linear classifier, and obtain the second sub-prediction vector corresponding to the data of each sub-image at different levels;

[0377] The second average aggregation processing unit 63 is used to perform average aggregation processing on the second sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the second attribute prediction vector corresponding to the data of each sub-image;

[0378] The third sub-prediction vector acquisition unit 62 is used to classify the second feature vectors corresponding to the data of each sub-image at different levels by using a non-linear classifier, and acquire the third sub-prediction vector corresponding to the data of each sub-image at different levels;

[0379] The third average aggregation processing unit 64 is used to perform average aggregation processing on the third sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the third attribute prediction vector corresponding to the data of each sub-image;

[0380] The prediction model building module 7 is also used to perform splicing processing on the intermediate prediction models corresponding to the data of each sub-image to establish a target attribute prediction model for predicting product attributes.

[0381] Wherein, when the intermediate prediction models corresponding to the data of each sub-image have equal weights, the intermediate prediction models corresponding to the data of each sub-image are spliced ​​to establish a target attribute prediction model for predicting product attributes.

[0382] The following is a specific example:

[0383] 1) Divide the image data

[0384] The image data is divided into M parts. Each time the model is trained, (M-1) parts are used as the training set, and the remaining 1 part is used as the validation set.

[0385] 2) Preprocess the image data

[0386] Perform data enhancement preprocessing on the image, such as changing color, flipping left and right, cropping, etc.;

[0387] As for image cropping, according to actual observation experience, the distribution of the center position of the products in the images of most products on the Internet platform has a strong regularity, such as being concentrated near the center of the image or the vertical center line, and the size of the products does not vary greatly; therefore, the following random cropping scheme can be used to crop the image: scale the width and height of the image to a set size (such as 224 pixels), and then construct an 8×8 regular grid on the image (such as Figure 4 As shown in the figure, three square areas with central side lengths of 56, 112, and 168 (represented by a, b, and c, respectively) are taken as random cropping centers, and the cropping sizes are 168, 112, and 56, respectively. The horizontal positions are sampled according to the normal distribution (that is, the image center is the mean and the variance is σ), and the vertical positions are sampled according to the uniform distribution. This sampling method can effectively integrate the "attention" prior distribution. In addition, for each type of cropping size, the four corners of the square and the 7 points intersecting the vertical center axis of the image are retained as fixed sampling points, so that a total of 27 fixed cropping methods are used for both training and testing.

[0388] 3) Establish basic model and basic model group

[0389] The data of each sub-image is processed at different levels including:

[0390] (1) Use a deep residual network (such as Resnetv2) to perform convolution processing on the data of sub-images at different levels, obtain N residual modules, and obtain the first attribute prediction vector output by the residual network, where N is a positive integer;

[0391] Perform GlobalAvgPool (global average pooling) on ​​the Nth residual module to obtain the intermediate feature vector;

[0392] The Softmax regression algorithm is used to regress the intermediate feature vector to obtain the first attribute prediction vector;

[0393] (2) The first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process, and the first feature vector is subjected to global average pooling processing to obtain a second feature vector;

[0394] PCA (principal component analysis) is used to perform dimensionality reduction processing on the second eigenvector to obtain the second eigenvector after dimensionality reduction processing;

[0395] Among them, some features were retained through principal component analysis technology, and the variance was retained at 95%.

[0396] A linear model SVM (support vector machine) is used to classify the second feature vector after dimension reduction to obtain a second attribute prediction vector;

[0397] The nonlinear model CART (regression classification tree) is used to classify the second eigenvector after dimensionality reduction to obtain the third attribute prediction vector.

[0398] The formula corresponding to the first attribute prediction vector is as follows:

[0399]

[0400] The product has m attributes, w i represents the first training parameter, w j represents the second training parameter, x represents different attributes, P i Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute, i, j = 0, 1, ..., m and are all integers;

[0401] The formula corresponding to the second attribute prediction vector is as follows:

[0402] y ij =w ij x+b ij ,

[0403] w ij represents the third training parameter, b ij represents the fourth training parameter, y ij Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute;

[0404] The formula corresponding to the third attribute prediction vector is as follows:

[0405]

[0406] N i Indicates that N j Indicates that P i Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute;

[0407] Among them, the first attribute prediction vector obtained by the Softmax layer of the deep residual network Resnetv2 is used as the Softmax main classifier, the SVM auxiliary classifier of the second attribute prediction vector obtained by SVM and the third attribute prediction vector obtained by CART are used as CART auxiliary classifiers, and then the main classifier and the auxiliary classifier are combined to obtain the final classification result, thereby effectively improving the classification performance based on a single model.

[0408] (3) For different levels of data of each sub-image, corresponding basic models are constructed respectively according to the above steps (1) and (2), and a basic model group corresponding to the data of each sub-image is constructed based on the basic model.

[0409] For example, if the six levels of data for each sub-image are trained separately, then the data for each sub-image corresponds to six basic models, which together form a basic model group, that is, the data for each sub-image corresponds to a basic model group. The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process, and the more abstract the information contained, thereby capturing the multi-layer semantic knowledge contained in the image and learning more image data. Generally, 3-8 residual modules are used.

[0410] Each basic model contains three classifiers, namely, Softmax main classifier, SVM auxiliary classifier and CART auxiliary classifier, so each basic model group includes 18 classifiers, that is, the data of each sub-image corresponds to 18 classifiers.

[0411] For a basic model group, each basic model outputs the first sub-prediction vector corresponding to the Softmax main classifier, the second sub-prediction vector corresponding to the SVM auxiliary classifier, and the third sub-prediction vector corresponding to the CART auxiliary classifier; then the first sub-prediction vectors of different basic models are averaged and aggregated to obtain the first attribute prediction vector corresponding to the sub-image data; the second sub-prediction vectors of different basic models are averaged and aggregated to obtain the second attribute prediction vector corresponding to the sub-image data; the third sub-prediction vectors of different basic models are averaged and aggregated to obtain the third attribute prediction vector corresponding to the sub-image data, so as to obtain three attribute prediction vectors corresponding to a basic model group, and the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector are spliced ​​to obtain an intermediate prediction model corresponding to a basic model group; finally, the intermediate prediction models corresponding to each basic model group are spliced ​​to establish a target attribute prediction model for predicting product attributes, thereby predicting product attributes.

[0412] In this embodiment, on the one hand, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as a basic model, and a basic model group is used to replace the basic model used in the usual solution to better capture the multi-layer semantic knowledge contained in the image. Each basic model works at a semantic level, and the entire basic model group works at a multi-layer semantic level, thereby learning more image data and having stronger prediction capabilities; on the other hand, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually merge, thereby improving nonlinear capabilities, reducing feature dimensions, and improving overall learning efficiency, thereby overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities.

[0413] Example 9

[0414] like Fig. 9 As shown, the product attribute prediction system of this embodiment is a further improvement of Embodiment 7, specifically:

[0415] After using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image, the first prediction vector acquisition module 3 includes a first sub-prediction vector acquisition unit 31 and a first regression processing unit 33;

[0416] The first sub-prediction vector acquisition unit 31 is used to acquire the first sub-prediction vector corresponding to the data of each sub-image at different levels;

[0417] The first regression processing unit 33 is used to use a regression algorithm to perform regression processing on the first sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the first attribute prediction vector corresponding to the first image data;

[0418] The intermediate prediction vector acquisition module 6 includes a second sub-prediction vector acquisition unit 61, a third sub-prediction vector acquisition unit 63, a first regression processing unit 65 and a second regression processing unit 66;

[0419] The second sub-prediction vector acquisition unit 61 is used to classify the second feature vectors corresponding to the data of each sub-image at different levels by using a linear classifier, and obtain the second sub-prediction vector corresponding to the data of each sub-image at different levels;

[0420] The first regression processing unit 65 is used to use a regression algorithm to perform regression processing on the second sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the second attribute prediction vector corresponding to the first image data;

[0421] The third sub-prediction vector acquisition unit 63 is used to classify the second feature vectors corresponding to the data of each sub-image at different levels by using a non-linear classifier, and acquire the third sub-prediction vector corresponding to the data of each sub-image at different levels;

[0422] The second regression processing unit 66 is used to perform regression processing on the third sub-prediction vectors corresponding to the data of different sub-images at different levels by using a regression algorithm, so as to obtain the third attribute prediction vector corresponding to the first image data.

[0423] The following is a specific example:

[0424] 1) Divide the image data

[0425] The image data is divided into M parts. Each time the model is trained, (M-1) parts are used as the training set, and the remaining 1 part is used as the validation set.

[0426] 2) Preprocess the image data

[0427] Perform data enhancement preprocessing on the image, such as changing color, flipping left and right, cropping, etc.;

[0428] As for image cropping, according to actual observation experience, the distribution of the center position of the products in the images of most products on the Internet platform has a strong regularity, such as being concentrated near the center of the image or the vertical center line, and the size of the products does not vary greatly; therefore, the following random cropping scheme can be used to crop the image: scale the width and height of the image to a set size (such as 224 pixels), and then construct an 8×8 regular grid on the image (such as Figure 4 As shown in the figure, three square areas with central side lengths of 56, 112, and 168 (represented by a, b, and c, respectively) are taken as random cropping centers, and the cropping sizes are 168, 112, and 56, respectively. The horizontal positions are sampled according to the normal distribution (that is, the image center is the mean and the variance is σ), and the vertical positions are sampled according to the uniform distribution. This sampling method can effectively integrate the "attention" prior distribution. In addition, for each type of cropping size, the four corners of the square and the 7 points intersecting the vertical center axis of the image are retained as fixed sampling points, so that a total of 27 fixed cropping methods are used for both training and testing.

[0429] 3) Establish basic model and basic model group

[0430] The data of each sub-image is processed at different levels including:

[0431] (1) Use a deep residual network (such as Resnetv2) to perform convolution processing on the data of sub-images at different levels, obtain N residual modules, and obtain the first attribute prediction vector output by the residual network, where N is a positive integer;

[0432] Perform GlobalAvgPool (global average pooling) on ​​the Nth residual module to obtain the intermediate feature vector;

[0433] The Softmax regression algorithm is used to regress the intermediate feature vector to obtain the first attribute prediction vector;

[0434] (2) The first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process, and the first feature vector is subjected to global average pooling processing to obtain a second feature vector;

[0435] PCA (principal component analysis) is used to perform dimensionality reduction processing on the second eigenvector to obtain the second eigenvector after dimensionality reduction processing;

[0436] Among them, some features were retained through principal component analysis technology, and the variance was retained at 95%.

[0437] A linear model SVM (support vector machine) is used to classify the second feature vector after dimension reduction to obtain a second attribute prediction vector;

[0438] The nonlinear model CART (regression classification tree) is used to classify the second eigenvector after dimensionality reduction to obtain the third attribute prediction vector.

[0439] The formula corresponding to the first attribute prediction vector is as follows:

[0440]

[0441] The product has m attributes, w i represents the first training parameter, w j represents the second training parameter, x represents different attributes, P i Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute, i, j = 0, 1, ..., m and are all integers;

[0442] The formula corresponding to the second attribute prediction vector is as follows:

[0443] y ij =w ij x+b ij ,

[0444] w ij represents the third training parameter, b ijrepresents the fourth training parameter, y ij Represents the attribute prediction vector composed of the predicted values ​​corresponding to each attribute;

[0445] The formula corresponding to the third attribute prediction vector is as follows:

[0446]

[0447] N i Indicates the number of samples of class i in the leaf node, N j represents the number of samples of class j in the leaf node, P i Represents the prediction vector corresponding to class i (i.e., the probability corresponding to class i);

[0448] Among them, the first attribute prediction vector obtained by the Softmax layer of the deep residual network Resnetv2 is used as the Softmax main classifier, the SVM auxiliary classifier of the second attribute prediction vector obtained by SVM and the third attribute prediction vector obtained by CART are used as CART auxiliary classifiers, and then the main classifier and the auxiliary classifier are combined to obtain the final classification result, thereby effectively improving the classification performance based on a single model.

[0449] (3) For different levels of data of each sub-image, corresponding basic models are constructed respectively according to the above steps (1) and (2), and a basic model group corresponding to the data of each sub-image is constructed based on the basic model.

[0450] For example, if the six levels of data for each sub-image are trained separately, then the data for each sub-image corresponds to six basic models, which together form a basic model group, that is, the data for each sub-image corresponds to a basic model group. The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process, and the more abstract the information contained, thereby capturing the multi-layer semantic knowledge contained in the image and learning more image data. Generally, 3-8 residual modules are used.

[0451] Each basic model contains three classifiers, namely, Softmax main classifier, SVM auxiliary classifier and CART auxiliary classifier, so each basic model group includes 18 classifiers, that is, the data of each sub-image corresponds to 18 classifiers.

[0452] For a basic model group, each basic model outputs the first sub-prediction vector corresponding to the Softmax main classifier, the second sub-prediction vector corresponding to the SVM auxiliary classifier, and the third sub-prediction vector corresponding to the CART auxiliary classifier; then the first sub-prediction vectors in different basic model groups are regressed to obtain the first attribute prediction vector corresponding to the image data; the second sub-prediction vectors in different basic model groups are regressed to obtain the second attribute prediction vector corresponding to the image data; the third sub-prediction vectors in different basic model groups are regressed to obtain the third attribute prediction vector corresponding to the sub-image data, so as to obtain the three attribute prediction vectors corresponding to the image data, and the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector are spliced ​​to obtain, and a target attribute prediction model for predicting product attributes is established to further predict product attributes.

[0453] In this embodiment, on the one hand, a deep convolutional model is used as a feature extractor, and classifiers with different characteristics are mixed as a basic model, and a basic model group is used to replace the basic model used in the usual solution to better capture the multi-layer semantic knowledge contained in the image. Each basic model works at a semantic level, and the entire basic model group works at a multi-layer semantic level, thereby learning more image data and having stronger prediction capabilities; on the other hand, unlike the ordinary single-layer directly using an integration method, a two-layer integration method is used to gradually fuse heterogeneous models, thereby improving nonlinear capabilities, reducing feature dimensions, and improving overall learning efficiency, thereby overcoming the serious overfitting and high prediction errors of existing prediction methods, and improving the prediction accuracy of product attributes. The entire method has good robustness and practical application capabilities.

[0454] Example 10

[0455] like Fig.10 As shown, the product attribute prediction system of this embodiment is a further improvement of the embodiment 6, specifically:

[0456] The prediction system also includes a verification set acquisition module, which is also used to select the Mk molecular data set to form a verification set;

[0457] The data in the validation set are preprocessed, and a basic model group is trained on each training set, with a total of m basic model groups trained. Each basic model group contains 6 basic models, which have a similar structure, except that the number of residuals of the resnetv2 model is different, generally 3 to 8. The feature map outputs of the last two residual models are used to extract features as training samples for SVM and CART, and the model hyperparameters are selected through the validation set.

[0458] The validation set is enhanced using data preprocessing methods, and then m basic model groups are input respectively. Each basic model in each basic model group outputs three types of prediction vectors: Softmax, SVM, and CART. The outputs of all homogeneous models in the basic model group are averaged and aggregated, so that three prediction vectors are obtained as the feature representation of this model group.

[0459] Different model groups obtain different feature representations. Assuming that there are c categories of clothing attributes, the output feature dimension of each basic model group is [c+c(c-1) / 2+c], and the total feature representation dimension obtained by the basic model layer is [mc(c+3) / 2]; each intermediate classifier in the middle layer uses part of the features as input, and the input dimensions of each intermediate classifier are mc, [mc(c-1) / 2], mc, respectively, and the corresponding attributes are the same as the original input image.

[0460] Specifically, the output dimension of CART in the basic model is c, and the feature value is the distribution of c-type samples in the leaf node into which the sample falls.

[0461] For example, if the number of samples of class c contained in the leaf node is m1, m2, ..., m c , then the output prediction vector is (m1 / m, m2 / m, …, m c / m), where The output dimension of the SVM in the basic model is [c(c-1) / 2]. Since SVM is a binary classifier, in order to solve the c-class problem, it is necessary to train [c(c-1) / 2] classifiers, that is, c i c j A classifier is trained for every two types of samples, and the prediction value of each classifier is used as the feature value of that dimension; the softmax prediction value is used as the c-dimensional feature vector.

[0462] Therefore, the feature outputs of the homogeneous models in each basic model group are first averaged and aggregated, that is, the feature dimensions obtained after the average aggregation of the homogeneous model SVM, softmax, and CART outputs in the k basic models are c, c(c-1) / 2, c;

[0463] The output features of the homogeneous models in different basic model groups are concatenated. Assuming there are m basic model groups, the feature input dimensions of the intermediate classifier corresponding to each homogeneous model are mc, [mc(c-1) / 2], mc, and the output dimension is c. The input is a 3-category heterogeneous model with a dimension of 3c, and the output is a c-dimensional prediction probability vector;

[0464] All of the above classifiers use Softmax as the weight learner of the combined model, avoiding a simple voting mechanism and enhancing the predictive ability of the overall model.

[0465] The prediction system also includes an actual attribute acquisition module 11, a verification module 12 and a judgment module 13;

[0466] The actual attribute acquisition module 11 is used to acquire the actual attribute of the first product being trained;

[0467] The verification module 12 is used to input the sub-data set in the verification set into the target attribute prediction model to obtain the prediction result of the first product;

[0468] The judgment module 13 is used to judge whether the prediction result is consistent with the actual attribute of the trained product. If so, it is determined that the target attribute prediction model is feasible; otherwise, the target attribute prediction model is retrained until the prediction result is consistent with the actual attribute of the first trained product.

[0469] In this embodiment, the accuracy of the target attribute prediction model is verified through a validation set to ensure the accuracy of the target attribute prediction model in predicting product attributes, so that the entire method has good robustness and practical application capabilities.

[0470] Embodiment 11

[0471] Fig.11 This is a schematic diagram of the structure of an electronic device provided in Embodiment 11 of the present invention. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the program, the method for predicting product attributes in any one of Embodiments 1 to 5 is implemented. Fig.11 The electronic device 40 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0472] like Fig.11 As shown, the electronic device 40 may be in the form of a general-purpose computing device, for example, it may be a server device. The components of the electronic device 40 may include, but are not limited to: at least one processor 41, at least one memory 42, and a bus 43 connecting different system components (including the memory 42 and the processor 41).

[0473] The bus 43 includes a data bus, an address bus, and a control bus.

[0474] The memory 42 may include a volatile memory, such as a random access memory (RAM) 421 and / or a cache memory 422 , and may further include a read-only memory (ROM) 423 .

[0475] The memory 42 may also include a program / utility 325 having a set (at least one) of program modules 424, such program modules 424 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0476] The processor 41 executes various functional applications and data processing by running the computer programs stored in the memory 42, such as the method for predicting product attributes in any one of Embodiments 1 to 5 of the present invention.

[0477] The electronic device 40 may also communicate with one or more external devices 44 (e.g., keyboards, pointing devices, etc.). Such communication may be performed via an input / output (I / O) interface 45. Furthermore, the model generating device 40 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 46. Fig.11 As shown, the network adapter 46 communicates with other modules of the model-generated electronic device 40 via the bus 43. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the model-generated electronic device 40, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems, etc.

[0478] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to an embodiment of the present invention, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0479] Example 12

[0480] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the method for predicting product attributes in any one of Embodiments 1 to 5 are implemented.

[0481] The readable storage medium may include but is not limited to: a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device or any suitable combination of the above.

[0482] In a possible implementation, the present invention can also be implemented in the form of a program product, which includes program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps in the method for predicting product attributes in any one of embodiments 1 to 5.

[0483] Among them, the program code for executing the present invention can be written in any combination of one or more programming languages, and the program code can be executed completely on the user device, partially on the user device, as an independent software package, partially on the user device and partially on a remote device, or completely on the remote device.

[0484] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these are only examples, and the protection scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. A method for predicting product attributes, characterized in that: The prediction method comprises: Acquire first image data of a first product to be trained, and acquire a target training set according to the first image data; Performing attribute feature extraction processing on the target training set using a residual network, obtaining a first attribute prediction vector output by the residual network, and obtaining a first feature vector corresponding to a residual module obtained by the residual network during the attribute feature extraction process; Performing pooling processing on the first feature vector by using a pooling method to obtain a second feature vector; Using a linear classifier and a nonlinear classifier to classify the second feature vector to obtain an intermediate attribute prediction vector; The first attribute prediction vector and the intermediate attribute prediction vector are concatenated to establish a target attribute prediction model for predicting product attributes; Acquire second image data of a target product to be predicted; The second image data is input into the target attribute prediction model to predict the target attribute of the target product.

2. The method for predicting product attributes according to claim 1, characterized in that: The step of using a linear classifier and a nonlinear classifier to classify the second feature vector to obtain an intermediate attribute prediction vector comprises: Using a linear classifier to classify the second feature vector to obtain a second attribute prediction vector; Using a nonlinear classifier to classify the second feature vector to obtain a third attribute prediction vector; The step of concatenating the first attribute prediction vector and the intermediate attribute prediction vector to establish a target attribute prediction model for predicting product attributes comprises: The first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector are concatenated to establish the target attribute prediction model for predicting product attributes.

3. The method for predicting product attributes according to claim 1, characterized in that: After the step of performing pooling processing on the first feature vector by using a pooling method to obtain a second feature vector and before the step of performing classification processing on the second feature vector by using a linear classifier and a nonlinear classifier to obtain an intermediate attribute prediction vector, the method further includes: A principal component analysis technique is used to perform dimensionality reduction processing on the second eigenvector to obtain the second eigenvector after dimensionality reduction processing.

4. The method for predicting product attributes according to claim 1, characterized in that: The step of acquiring a target training set according to the first image data comprises: The first image data is divided into M sub-data sets, and k sub-data sets are selected to form the target training set, where k is less than M and both M and k are positive integers.

5. The method for predicting product attributes according to claim 2, characterized in that: The step of acquiring a target training set according to the first image data comprises: Divide the first image data into M sub-datasets, and select k of the sub-datasets to form a first training set, where k < M and M and k are both positive integers; The first training set is preprocessed by using at least one of color conversion processing, flip processing and cropping processing to obtain the preprocessed target training set.

6. The method for predicting product attributes according to claim 5, characterized in that: When the first training set is trimmed, data of a plurality of sub-images corresponding to the first image data are obtained; The data of each sub-image constitutes a target training set.

7. The method for predicting product attributes according to claim 5, characterized in that: When the first training set is trimmed, data of a plurality of sub-images corresponding to the first image data are obtained; The step of using a residual network to extract attribute features from the target training set includes: Using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image; The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process; The step of obtaining a first feature vector corresponding to the residual module obtained by the residual network in the attribute feature extraction process includes: Obtaining the first feature vector corresponding to the residual module obtained by the residual network in a process of extracting attribute features at different levels of data of each sub-image; The step of performing pooling processing on the first feature vector by using the pooling method to obtain the second feature vector comprises: The pooling method is used to perform pooling processing on the first feature vectors corresponding to the data of each sub-image at different levels, so as to obtain the second feature vectors corresponding to the data of each sub-image at different levels.

8. The method for predicting product attributes according to claim 7, characterized in that: After using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image, the step of obtaining the first attribute prediction vector includes: Obtaining a first sub-prediction vector corresponding to the data of each sub-image at a different level; Respectively perform average aggregation processing on the first sub-prediction vectors corresponding to the data of the same sub-image at different levels to obtain the first attribute prediction vector corresponding to the data of each sub-image; The step of using a linear classifier to classify the second feature vector to obtain a second attribute prediction vector comprises: Using the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels; Respectively perform average aggregation processing on the second sub-prediction vectors corresponding to the data of the same sub-image at different levels to obtain the second attribute prediction vector corresponding to the data of each sub-image; The step of using a nonlinear classifier to classify the second feature vector to obtain a third attribute prediction vector comprises: Using the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the third sub-prediction vectors corresponding to the data of each sub-image at different levels; Respectively perform average aggregation processing on the third sub-prediction vectors corresponding to the data of the same sub-image at different levels to obtain the third attribute prediction vector corresponding to the data of each sub-image; The step of concatenating the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector to establish the target attribute prediction model for predicting product attributes includes: Performing splicing processing on the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector to obtain an intermediate prediction model corresponding to the data of each of the sub-images; The intermediate prediction models corresponding to the data of each sub-image are spliced ​​to establish the target attribute prediction model for predicting product attributes.

9. The method for predicting product attributes according to claim 8, characterized in that: The step of performing splicing processing on the intermediate prediction models corresponding to the data of each sub-image to establish the target attribute prediction model for predicting product attributes comprises: In the case where the intermediate prediction models corresponding to the data of each sub-image have equal weights, the intermediate prediction models corresponding to the data of each sub-image are spliced ​​to establish the target attribute prediction model for predicting product attributes.

10. The method for predicting product attributes according to claim 7, characterized in that: After using the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image, the step of obtaining the first attribute prediction vector includes: Obtaining a first sub-prediction vector corresponding to the data of each sub-image at a different level; Using a regression algorithm to perform regression processing on the first sub-prediction vectors corresponding to the data of different sub-images at different levels, respectively, to obtain the first attribute prediction vector corresponding to the first image data; The step of using a linear classifier to classify the second feature vector to obtain a second attribute prediction vector comprises: Using the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels; Using the regression algorithm to perform regression processing on the second sub-prediction vectors corresponding to the data of different sub-images at different levels, respectively, to obtain the second attribute prediction vector corresponding to the first image data; The step of using a nonlinear classifier to classify the second feature vector to obtain a third attribute prediction vector comprises: Using the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, respectively, to obtain the third sub-prediction vectors corresponding to the data of each sub-image at different levels; The regression algorithm is used to perform regression processing on the third sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the third attribute prediction vector corresponding to the first image data.

11. The method for predicting product attributes according to claim 1, characterized in that: The step of using a residual network to extract attribute features from the target training set and obtaining a first attribute prediction vector output by the residual network comprises: Using a residual network to perform convolution processing on the target training set to obtain N residual modules, where N is a positive integer; Performing pooling processing on the Nth residual module using the pooling method to obtain an intermediate feature vector; The dimension of the intermediate feature vector is higher than the dimension of the first feature vector corresponding to the residual module; A regression algorithm is used to perform regression processing on the intermediate feature vector to obtain the first attribute prediction vector.

12. The method for predicting product attributes according to claim 11, characterized in that: When the residual network includes N residual modules during the attribute feature extraction process, the first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network during the attribute feature extraction process.

13. The method for predicting product attributes according to claim 10, characterized in that: The residual network includes a deep residual network or an Inception network; and / or, The linear classifier comprises a support vector machine; and / or, The nonlinear classifier includes a regression classification tree or a random forest algorithm; and / or, The pooling method includes global average pooling; and / or, The regression algorithm includes a multi-class logistic regression algorithm.

14. The method for predicting product attributes according to claim 1, characterized in that: The step of inputting the second image data into the target attribute prediction model to predict the target attribute of the target product comprises: Inputting the second image data into the target attribute prediction model, and selecting the maximum vector value in the prediction vector corresponding to the target attribute prediction model; The attribute corresponding to the maximum vector value is determined as the target attribute of the target product.

15. The method for predicting product attributes according to claim 4 or 5, characterized in that: The step of dividing the first image data into M sub-datasets and selecting k sub-datasets to form a first training set also includes: Select Mk sub-datasets to form a validation set; The step of obtaining the target attribute prediction model further includes: Obtaining actual attributes of the first product being trained; Inputting the sub-dataset in the validation set into the target attribute prediction model to obtain a prediction result of the first product; Determine whether the prediction result is consistent with the actual attribute of the trained product. If so, determine that the target attribute prediction model is feasible; otherwise, retrain the target attribute prediction model until the prediction result is consistent with the actual attribute of the first product being trained.

16. A product attribute prediction system, characterized in that: The prediction system also includes a first image data acquisition module, a target training set acquisition module, a first prediction vector acquisition module, a first feature vector acquisition module, a second feature vector acquisition module, an intermediate prediction vector acquisition module, a prediction model establishment module, a second image data acquisition module and a target attribute prediction module; The first image data acquisition module is used to acquire first image data of a first product to be trained; The target training set acquisition module is used to acquire a target training set according to the first image data; The first prediction vector acquisition module is used to use a residual network to perform attribute feature extraction processing on the target training set to obtain a first attribute prediction vector output by the residual network; The first feature vector acquisition module is used to obtain a first feature vector corresponding to the residual module obtained by the residual network during the attribute feature extraction process; The second feature vector acquisition module is used to perform pooling processing on the first feature vector by using a pooling method to obtain a second feature vector; The intermediate prediction vector acquisition module is used to classify the second feature vector using a linear classifier and a nonlinear classifier to obtain an intermediate attribute prediction vector; The prediction model building module is used to perform splicing processing on the first attribute prediction vector and the intermediate attribute prediction vector to establish a target attribute prediction model for predicting product attributes; The second image data acquisition module is used to acquire second image data of the target product to be predicted; The target attribute prediction module is used to input the second image data into the target attribute prediction model to predict the target attribute of the target product.

17. The product attribute prediction system according to claim 16, characterized in that: The intermediate prediction vector acquisition module is used to classify the second feature vector using a linear classifier to obtain a second attribute prediction vector; The intermediate prediction vector acquisition module is further used to classify the second feature vector using a nonlinear classifier to obtain a third attribute prediction vector; The prediction model building module is used to perform splicing processing on the first attribute prediction vector, the second attribute prediction vector and the third attribute prediction vector to establish the target attribute prediction model for predicting product attributes.

18. The product attribute prediction system according to claim 16, characterized in that: The prediction system also includes a principal component analysis module; The principal component analysis module is used to perform dimensionality reduction processing on the second eigenvector using the principal component analysis technology to obtain the second eigenvector after dimensionality reduction processing.

19. The product attribute prediction system according to claim 16, characterized in that: The target training set acquisition module is used to divide the first image data into M sub-data sets, and select k sub-data sets to form the target training set, k<M and M and k are both positive integers.

20. The product attribute prediction system according to claim 17, characterized in that: The target training set acquisition module includes a training set acquisition unit and a preprocessing unit; The training set acquisition unit is used to divide the first image data into M sub-data sets, and select k sub-data sets to form a first training set, where k < M and M and k are both positive integers; The preprocessing unit is used to preprocess the first training set by using at least one processing method of color conversion processing, flip processing and clipping processing to obtain the preprocessed target training set.

21. The product attribute prediction system according to claim 20, characterized in that: The preprocessing unit is used to perform trimming processing on the first training set to obtain data of multiple sub-images corresponding to the first image data; The data of each sub-image constitutes a target training set.

22. The product attribute prediction system according to claim 20, characterized in that: The preprocessing unit is used to perform trimming processing on the first training set to obtain data of multiple sub-images corresponding to the first image data; The first prediction vector acquisition module is further used to use the residual network to perform attribute feature extraction processing at different levels on the data of each sub-image; The deeper the level, the more residual modules the residual network corresponds to during the attribute feature extraction process; The first feature vector acquisition module is further used to acquire the first feature vector corresponding to the residual module obtained by the residual network in the process of extracting attribute features at different levels of data of each sub-image; The second feature vector acquisition module is also used to use the pooling method to perform pooling processing on the first feature vectors corresponding to the data of each sub-image at different levels, so as to obtain the second feature vectors corresponding to the data of each sub-image at different levels.

23. The product attribute prediction system according to claim 22, characterized in that: After the residual network is used to perform attribute feature extraction processing at different levels on the data of each of the sub-images, the first prediction vector acquisition module includes a first sub-prediction vector acquisition unit and a first average aggregation processing unit; The first sub-prediction vector acquisition unit is used to acquire the first sub-prediction vector corresponding to the data of each sub-image at different levels; The first average aggregation processing unit is used to perform average aggregation processing on the first sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the first attribute prediction vector corresponding to the data of each sub-image; The intermediate prediction vector acquisition module includes a second sub-prediction vector acquisition unit, a third sub-prediction vector acquisition unit, a second average aggregation processing unit and a third average aggregation processing unit; The second sub-prediction vector acquisition unit is used to use the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels; The second average aggregation processing unit is used to perform average aggregation processing on the second sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the second attribute prediction vector corresponding to the data of each sub-image; The third sub-prediction vector acquisition unit is used to use the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the third sub-prediction vector corresponding to the data of each sub-image at different levels; The third average aggregation processing unit is used to perform average aggregation processing on the third sub-prediction vectors corresponding to the data of the same sub-image at different levels, to obtain the third attribute prediction vector corresponding to the data of each sub-image; The prediction model building module is also used to perform splicing processing on the intermediate prediction models corresponding to the data of each sub-image to establish the target attribute prediction model for predicting product attributes.

24. The product attribute prediction system according to claim 23, characterized in that: The prediction model establishment module is also used to splice the intermediate prediction models corresponding to the data of each sub-image when the intermediate prediction models corresponding to the data of each sub-image have equal weights, so as to establish the target attribute prediction model for predicting product attributes.

25. The product attribute prediction system according to claim 22, characterized in that: After the residual network is used to perform attribute feature extraction processing at different levels on the data of each sub-image, the first prediction vector acquisition module includes a first sub-prediction vector acquisition unit and a first regression processing unit; The first sub-prediction vector acquisition unit is used to acquire the first sub-prediction vector corresponding to the data of each sub-image at different levels; The first regression processing unit is used to use a regression algorithm to perform regression processing on the first sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the first attribute prediction vector corresponding to the first image data; The intermediate prediction vector acquisition module includes a second sub-prediction vector acquisition unit, a third sub-prediction vector acquisition unit, a first regression processing unit and a second regression processing unit; The second sub-prediction vector acquisition unit is used to use the linear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the second sub-prediction vectors corresponding to the data of each sub-image at different levels; The first regression processing unit is used to use the regression algorithm to perform regression processing on the second sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the second attribute prediction vector corresponding to the first image data; The third sub-prediction vector acquisition unit is used to use the nonlinear classifier to classify the second feature vectors corresponding to the data of each sub-image at different levels, and obtain the third sub-prediction vector corresponding to the data of each sub-image at different levels; The second regression processing unit is used to use the regression algorithm to perform regression processing on the third sub-prediction vectors corresponding to the data of different sub-images at different levels, so as to obtain the third attribute prediction vector corresponding to the first image data.

26. The product attribute prediction system according to claim 16, characterized in that: The first prediction vector acquisition module includes a convolution processing unit, a pooling processing unit and a third regression processing unit; The convolution processing unit is used to perform convolution processing on the target training set using a residual network to obtain N residual modules, where N is a positive integer; The pooling processing unit is used to perform pooling processing on the Nth residual module using the pooling method to obtain an intermediate feature vector; The dimension of the intermediate feature vector is higher than the dimension of the first feature vector corresponding to the residual module; The third regression processing unit is used to perform regression processing on the intermediate feature vector using a regression algorithm to obtain the first attribute prediction vector.

27. The product attribute prediction system according to claim 26, characterized in that: When the first prediction vector acquisition module uses a residual network to include N residual modules in the attribute feature extraction process, the first feature vector includes feature vectors corresponding to the N-1th residual module and the Nth residual module respectively obtained by the residual network in the attribute feature extraction process.

28. The product attribute prediction system according to claim 25, characterized in that: The residual network includes a deep residual network or an Inception network; and / or, The linear classifier comprises a support vector machine; and / or, The nonlinear classifier includes a regression classification tree or a random forest algorithm; and / or, The pooling method includes global average pooling; and / or, The regression algorithm includes a multi-class logistic regression algorithm.

29. The product attribute prediction system according to claim 16, characterized in that: The target attribute prediction module includes a selection unit and a determination unit; The selection unit is used to input the second image data into the target attribute prediction model, and select the maximum vector value in the prediction vector corresponding to the target attribute prediction model; The determining unit is used to determine that the attribute corresponding to the maximum vector value is the target attribute of the target product.

30. The product attribute prediction system according to claim 19 or 20, characterized in that: The prediction system also includes a verification set acquisition module; The verification set acquisition module is used to select Mk pieces of the sub-datasets to form a verification set; The prediction system also includes an actual attribute acquisition module, a verification module and a judgment module; The actual attribute acquisition module is used to acquire the actual attribute of the first product being trained; The verification module is used to input the sub-dataset in the verification set into the target attribute prediction model to obtain the prediction result of the first product; The judgment module is used to judge whether the prediction result is consistent with the actual attribute of the trained product, and if so, determine that the target attribute prediction model is feasible; Otherwise, the target attribute prediction model is retrained until the prediction result is consistent with the actual attribute of the first product being trained.

31. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for predicting product attributes according to any one of claims 1 to 15 is implemented.

32. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting product attributes according to any one of claims 1 to 15 are implemented.

Images