A method and system for intelligent painting scoring based on attention convolutional network

Abstract

The present invention discloses a method and system for intelligently scoring paintings based on an attention convolutional network. The method comprises: constructing a coarse classifier and a fine classifier based on the attention convolutional network; constructing an intelligent scoring framework based on the coarse and fine classifiers; classifying the collected paintings according to scores to obtain input data, inputting the input data set into the coarse classifier, and then inputting the output of the coarse classifier into the fine classifier to obtain classification information; assigning a score to the classification information to obtain a final score. The system comprises: a data set acquisition module, a network construction module, a framework construction module, and an assignment scoring module. By using the present invention, problems caused by pre-trained networks can be circumvented, and paintings can be scored objectively, accurately, and quickly. As a method and system for intelligently scoring paintings based on an attention convolutional network, the present invention can be widely used in the field of intelligent image scoring.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent image scoring, and in particular to a method and system for intelligent painting scoring based on an attention convolutional network. Background Art

[0002] Grading paintings is an important part of art exams and an essential part of daily art teaching. Currently, this work is mainly handled by art teachers, who grade paintings manually. However, manual grading has problems such as low scoring efficiency, lack of objectivity and fairness, and a large consumption of art teachers. Therefore, with the deep integration of the Internet and the education industry, many painting grading methods based on machine learning and artificial intelligence have emerged. Most of these painting grading methods based on machine learning and artificial intelligence use pre-trained networks to extract painting features. Due to the characteristics of the pre-trained networks themselves, painting grading methods that use pre-trained networks to extract painting features have the following problems:

[0003] 1) Pre-trained networks are generally large networks, which take up a lot of resources and take a long time to infer when extracting image features, resulting in slow scoring.

[0004] 2) Pre-trained models are typically pre-trained on the Imagenet dataset, which contains images from a wide range of fields, in addition to some art. The features extracted directly from these images using the pre-trained model are not representative, leading to significant discrepancies between the model's predicted results and the actual scores.

[0005] 3) Pre-trained networks are usually composed of multi-layer convolutional neural networks, which have great advantages in extracting local features of images. However, fine art paintings occupy a larger scoring weight in terms of global relationships. Scoring only based on local features extracted by pre-trained networks ignores the importance of global features in paintings, resulting in overly one-sided scoring. Summary of the Invention

[0006] In order to solve the above technical problems, the goal of the present invention is to provide a method and system for intelligent painting scoring based on an attention convolutional network, which can avoid the problems caused by pre-trained networks and score paintings objectively, accurately and quickly.

[0007] The first technical solution adopted by the present invention is: a method for intelligently scoring paintings based on an attention convolutional network, comprising the following steps:

[0008] Classify the collected paintings according to their levels to obtain the input data set;

[0009] A classification network is constructed based on global features and local features to obtain an attention convolutional network;

[0010] Based on the attention convolutional network, a coarse classifier and a fine classifier are constructed to obtain an intelligent scoring framework;

[0011] Input the input data set into the intelligent scoring framework to obtain the final score;

[0012] The coarse classifier and the fine classifier are trained respectively to obtain the trained coarse classifier and the fine classifier.

[0013] Furthermore, by classifying the collected paintings according to levels to obtain the input data set, this preferred step unifies the sizes and categories of the pictures, facilitating the subsequent effective training of the coarse classifier and the fine classifier.

[0014] Furthermore, the step of constructing a classification network based on global features and local features to obtain an attention convolutional network specifically includes:

[0015] Build an image embedding layer based on a two-dimensional convolutional neural network;

[0016] Construct a local feature extraction layer based on depthwise separable convolutional blocks;

[0017] Construct a global feature learning layer based on the Reshape function and multi-head attention mechanism module;

[0018] Construct a fine-grained classification layer based on the global average pooling function and nonlinear activation function;

[0019] The image embedding layer, local feature extraction layer, global feature learning layer and fine-grained classification layer are fitted to obtain the attention convolutional network.

[0020] Through this preferred step, the constructed attention convolutional network can effectively extract the local features of the work. At the same time, the introduction of a multi-head attention mechanism module is used to focus on the global object relationship of color still life paintings, which is conducive to improving the classification accuracy.

[0021] Furthermore, the step of constructing a coarse classifier and a fine classifier based on the attention convolutional network to obtain an intelligent scoring framework specifically includes:

[0022] Based on the attention convolutional network, the depth-wise separable convolutional block is taken as the backbone network of the model to obtain a coarse classifier;

[0023] Based on the attention convolutional network, the depth-wise separable convolutional block is used as the backbone network of the model to obtain a detailed classifier.

[0024] Construct a scoring module based on random mood coefficient and assignment function;

[0025] The coarse classifier, fine classifier and scoring module are fitted to obtain an intelligent scoring framework.

[0026] Through this optimization step, the constructed intelligent scoring framework is divided into three stages: coarse classification, fine classification, and basic score assignment and scoring. Each stage focuses on the corresponding task, which improves the classification accuracy.

[0027] Furthermore, the step of inputting the input data set into the intelligent scoring framework to obtain the final score specifically includes:

[0028] The image embedding layer based on the coarse classifier projects the input data set to obtain the multi-channel first-class feature space projection;

[0029] Applying the local feature extraction layer of the coarse classifier to extract features from the first-category feature space projection of the multi-channel to obtain the first-category high-dimensional local features;

[0030] The global feature learning layer of the coarse classifier is used to supplement the learning of the first type of high-dimensional local features to obtain the interactive information between the first type of local features and the first type of global features;

[0031] Applying the fine-grained classification layer of the coarse classifier to classify the interactive information of the first-category local features and the first-category global features to obtain the first-category classification information;

[0032] The image embedding layer based on the detailed classifier projects the first category classification information to obtain a multi-channel second category feature space projection;

[0033] Applying the local feature extraction layer of the fine classifier to extract features from the second-category feature space projection of the multi-channel to obtain the second-category high-dimensional local features;

[0034] The global feature learning layer of the detailed classifier is used to supplement the learning of the second type of high-dimensional local features to obtain the interactive information between the second type of local features and the second type of global features;

[0035] Applying the fine-grained classification layer of the fine classifier to classify the interaction information of the second type of local features and the second type of global features to obtain the second type of classification information;

[0036] The scoring module is applied to assign scores to the second category of classification information to obtain the final score.

[0037] Through this optimization step, color still life paintings are classified from coarse to fine and then assigned specific score values. The mood factor is used to simulate the uncertainty of the actual scoring process, reducing the scoring error and making the final score closer to the actual estimated scoring result.

[0038] Furthermore, the step of training the coarse classifier and the fine classifier respectively to obtain the trained coarse classifier and the fine classifier specifically includes:

[0039] Perform image enhancement on the input data set to obtain an enhanced input data set;

[0040] Divide the enhanced input dataset into a high- and low-class classification task to obtain a warm-up dataset;

[0041] Pre-train the attention convolutional network model based on the preheated dataset to obtain the initial weights;

[0042] The coarse classifier and the fine classifier are trained with gradient update based on the initialized model parameters of the initial weights and the remaining input data set to obtain the estimated scoring results.

[0043] Through this preferred step, the input data set is enhanced to avoid overfitting of the network during training, providing a better initialization weight, which is conducive to accelerating the convergence of the model. The performance of the coarse classifier and the fine classifier can also be evaluated by comparing the estimated scoring results with the actual scoring.

[0044] The second technical solution adopted by the present invention is: an intelligent painting scoring system based on an attention convolutional network, comprising:

[0045] The dataset acquisition module is used to classify the collected paintings according to their levels to obtain the input dataset;

[0046] The network construction module builds a classification network based on global features and local features to obtain an attention convolutional network;

[0047] The framework construction module builds coarse and fine classifiers based on the attention convolutional network to obtain an intelligent scoring framework;

[0048] The assignment scoring module is used to input the input data set into the intelligent scoring framework to obtain the final score.

[0049] The beneficial effects of the method and system of the present invention are as follows: the present invention takes into account the problems existing in the extraction of painting features by pre-trained networks, adopts an attention convolutional network to extract local and global features of the paintings, and constructs an intelligent scoring framework based on the attention convolutional network model. Preheating training is used to accelerate model convergence, which can avoid the problems caused by the pre-trained network and objectively, accurately and quickly score the paintings. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flowchart of the steps of an intelligent painting scoring method based on an attention convolutional network of the present invention;

[0051] Figure 2 This is a structural diagram of an intelligent painting scoring system based on an attention convolutional network in the present invention;

[0052] Figure 3 It is a set of input data set artwork images for an intelligent painting scoring method based on an attention convolutional network in the present invention;

[0053] Figure 4 This is a structural diagram of an attention convolutional network for an intelligent painting scoring method based on an attention convolutional network according to the present invention;

[0054] Figure 5 This is a visualization of the depth-separable convolutional blocks of the intelligent painting scoring method based on the attention convolutional network of the present invention;

[0055] Figure 6 This is a diagram of the coarse classifier structure of an intelligent painting scoring method based on an attention convolutional network in the present invention;

[0056] Figure 7 This is a structural diagram of a detailed classifier of an intelligent painting scoring method based on an attention convolutional network according to the present invention;

[0057] Figure 8 This is a diagram of the intelligent scoring framework of an intelligent painting scoring method based on an attention convolutional network in the present invention;

[0058] Figure 9 This is a network training flowchart of the intelligent painting scoring method based on the attention convolutional network of the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are provided for ease of description only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted based on the understanding of those skilled in the art.

[0060] Reference Figure 1 The present invention provides a method for intelligently scoring paintings based on an attention convolutional network, which includes the following steps:

[0061] S1. Collect a batch of color still life paintings of middle school students in art exams as input data sets for model training and testing. In order to ensure the practicality of the classifier training results, the collected training set contains at least 8,000 works, and each work needs to have a score given by a professional art teacher. Since the present invention is a global feature and a local feature, it is necessary to first classify the collected pictures according to different score segments, which are divided into 8 score levels in total: less than 60 points, 60-64 points, 65-69 points, 70-74 points, 75-79 points, 80-84 points, 85-89 points and 90 points and above. Secondly, in order to facilitate the subsequent effective training of the network, the size of the collected pictures will be uniformly adjusted to 367×256×3. As Figure 3 The figure shows student works from two exams in different score ranges, where each row represents works from the same exam.

[0062] S2. The task of fine-grained classification of color still life paintings is different from other fine-grained global features and local features. The classification of color still lifes should not only focus on the local details of the works, but also on the relationship and coordination between the main objects and secondary objects in the paintings. Therefore, the present invention builds an attention convolutional neural network based on global features and local features. Figure 4 As shown in FIG, the attention convolutional network includes an image embedding layer, a local feature extraction layer, a global feature learning layer, and a fine-grained classification layer.

[0063] S2.1. In the initial stage of the attention convolutional network, the input image first passes through an image embedding layer. The present invention uses a two-dimensional convolutional neural network as the image embedding layer for color still life paintings. The main purpose is to project it into a multi-channel feature space through two-dimensional convolution, preparing for the subsequent extraction of painting features.

[0064] S2.2, such as Figure 5 As shown in the figure, in the local feature extraction stage of the attention convolutional network, the depth-wise separable convolution proposed in this invention is different from the traditional convolution layer. First, the input feature map is projected into a new feature space through a 1×1 point-by-point convolution, channel information is integrated, and cross-channel correlation is learned. Then, depth convolution is performed on each new feature plane from the spatial dimension, and finally, the new content of the feature plane is integrated through a 1×1 point-by-point convolution. In this way, the channel correlation and spatial correlation of the convolutional neural network are fully decoupled, which can effectively improve the performance of the network while reducing the number of network parameters. Its formula is as follows:

[0065]

[0066]

[0067] Where Conv represents ordinary convolution, Y represents the output of the depth-wise separable convolution block, DWconv represents separable convolution, and B i Indicates the bias of each layer.

[0068] S2.3. After completing the local feature extraction in S2.2, in order to strengthen the association between global features and make the network pay attention to more information, the present invention designs a global feature learning layer. This layer, as a supplement to the depth-separable convolution block, can learn the features captured by the depth-separable convolution block and emphasize the potential global representation. The global feature learning layer consists of a multi-head attention module and a Reshape function in the Transformer framework. Since the input tensor of the multi-head attention module is two-dimensional, the local features in S2.2 need to be flattened using the Reshape function before being input into the multi-head attention module. The formula is as follows:

[0069] Attention(Q,K,V)=softmax(QK T )V

[0070] Where Q, K, and V are the values obtained by different linear projections of the same input X.

[0071] Assuming the number of heads is h, we need to downsample the dimension of input X by h times and execute h self-attention modules respectively to obtain the output value of low-dimensional space. i Corresponding to the output of a self-attention mechanism. They will be concatenated and re-projected to obtain the final value. This will help the network focus on information from multiple subspaces and enhance the network's ability to extract global features. The expression of the self-attention mechanism of the multi-head attention module is as follows:

[0072] MultiHead=Concat(Head1,Head2,…,Head h )

[0073] S2.4. After passing through the multi-head attention mechanism module, relevant key information that is helpful for classification is extracted and input into a fine-grained classification layer, which outputs the classification probability of each fine category. The category with the highest probability is the class ultimately recognized by the network. The fine-grained classification layer consists of a global average pooling function and a nonlinear softmax activation function. Its formula is as follows:

[0074]

[0075]

[0076] where p jRepresents the probability of the jth category, GHF represents the global average pooling output of local features and global features, represents the interactive information containing local features and global features, d represents the number of heads of the multi-head attention module, i represents the i-th depth-separable convolution block, n represents the number of depth-separable convolution blocks, and y j Represents the feature information of the j-th category.

[0077] S2.5. Fit the image embedding layer, local feature extraction layer, global feature learning layer and fine-grained classification layer in the order of image embedding layer first, then local feature extraction layer, followed by global feature learning layer, and finally fine-grained classification layer to obtain the attention convolutional network.

[0078] S3. Based on the attention convolutional network, the present invention proposes an intelligent scoring framework for classifying and scoring color still life works from coarse to fine. The proposed intelligent scoring framework is shown in Figure 8. Specifically, the scoring framework can be divided into the following three stages.

[0079] S3.1, such as Figure 6 As shown in the figure, to achieve the classification of color still life works, three depthwise separable convolutional blocks are taken as the backbone network of the model based on the attention convolutional network, resulting in a coarse classifier. The coarse classifier divides the results into four levels: above 80 points, 70-79 points, 60-69 points, and below 60 points.

[0080] S3.2, such as Figure 7 As shown in the figure, based on the complexity of the task, to achieve the classification of color still life paintings, two depthwise separable convolutional blocks are taken from the attention convolutional network as the backbone network of the model, resulting in a fine-scale classifier. Based on the results of the coarse classifier, fine-scale classifiers 1, 2, and 3 are constructed. Specifically, works with a score of 80 or above are input to fine-scale classifier 1, which classifies them into three categories: 80-84, 85-89, and 90 or above. Works with a score of 70-79 are input to fine-scale classifier 2, which classifies them into two categories: 70-74 and 75-79. Works with a score of 60-69 are input to fine-scale classifier 3, which classifies them into two categories: 60-64 and 65-69. After the operation of the coarse classifier and the fine classifier, the works can be divided into the following eight levels: below 60 points, 60-64 points, 65-69 points, 70-74 points, 75-79 points, 80-84 points, 85-89 points and above 90 points.

[0081] S3.3. To achieve accurate scoring of color still life paintings, a base score is assigned to each score level, and a random mood coefficient α is generated between 0 and 5. A value-assigned scoring module is constructed to simulate the uncertainty factors when teachers score, achieving more realistic and refined scoring. The final score of the work is composed of the base score and the random mood coefficient, which is expressed as follows:

[0082] Score(X)=Basescore(CCfilter(X))+α

[0083] Where Basecore represents the basic score assignment function, CCfliter represents the fine-grained classification function from coarse to fine, α (0≤α<5) is the randomly generated mood factor, and X is the input color still life work.

[0084] S3.4. The coarse classifier, fine classifier and scoring module are first fitted with a coarse classifier, then the fine classifier is fitted based on each coarse classification result, and finally the scoring module is fitted based on each fine classifier result to obtain an intelligent scoring framework.

[0085] S4. Input the input data set into the intelligent scoring framework to obtain the final score.

[0086] S4.1. Input the input data set into the intelligent scoring framework. First, the input data set reaches the image embedding layer of the coarse classifier. The image embedding layer of the coarse classifier is applied to perform image embedding on the input data set to obtain the first-class feature space projection of the multi-channel; then the local feature extraction layer of the coarse classifier is applied to perform feature extraction on the first-class feature space projection of the multi-channel to obtain the first-class high-dimensional local features; then the global feature learning layer of the coarse classifier is applied to perform supplementary learning on the first-class high-dimensional local features to obtain the interactive information of the first-class local features and the first-class global features; finally, the fine-grained classification layer of the coarse classifier is applied to classify the interactive information of the first-class local features and the first-class global features, and the first-class classification information is obtained from the output of the coarse classifier.

[0087] S4.2. After the first-category classification information is output from the coarse classifier, it reaches the image embedding layer of the fine classifier. The image embedding layer of the fine classifier is used to perform image embedding on the first-category classification information to obtain the second-category feature space projection of the multi-channel; then the local feature extraction layer of the fine classifier is used to perform feature extraction on the second-category feature space projection of the multi-channel to obtain the second-category high-dimensional local features; then the global feature learning layer of the fine classifier is used to perform supplementary learning on the second-category high-dimensional local features to obtain the interactive information between the second-category local features and the second-category global features; finally, the fine-grained classification layer of the fine classifier is used to classify the interactive information between the second-category local features and the second-category global features to obtain the second-category classification information.

[0088] S4.3. Assign a basic value to the second category of classification information to obtain a basic score;

[0089] S4.4. Add a random mood factor to the basic score to obtain the final score.

[0090] S5, using the training strategy to train the coarse classifier, fine classifier 1, fine classifier 2 and fine classifier 3 in the previous step respectively, and obtain the estimated scoring results; the training strategy proposed in the present invention includes: image enhancement stage, model warm-up stage and model training stage, the overall training process is as follows Figure 9 shown.

[0091] S5.1. In the image enhancement stage, to prevent the network from overfitting during training, the present invention uses an image flipping and small-angle rotation enhancement strategy to enhance the input dataset and obtain an enhanced dataset.

[0092] S5.2. In the model preheating stage, the present invention first divides a high- and low-classification task in the enhanced data set to obtain a preheating data set, and preheats the attention convolutional network model based on the preheating data set to obtain W j is the weight value of the jth round in the model warm-up phase. The model pre-training updates the weight according to the weight update rule. After m1 rounds of training, the optimal weight is W * , the weight update rule expression is as follows:

[0093]

[0094] Where α is the learning rate at this stage.

[0095] S5.3, take V j is the weight value of the jth round in the model training phase, and the best weight W obtained in the warm-up phase training is * Initialize the model parameters, then perform gradient update training according to the weight update rule. When the maximum number of training rounds m2 is reached, stop and output the estimated results of the training score. The initialization expression and weight update rule expression are as follows:

[0096]

[0097]

[0098] Where β represents the learning rate of this stage.

[0099] S5.4, according to Figure 9 The process in , use the collected data sets to train Figure 8The estimated scoring results obtained by using the coarse and fine classifiers in the algorithm show that the accuracy of each classifier has reached over 80%, and the multi-head attention mechanism plays a key role in this. Secondly, the present invention deploys the trained intelligent scoring framework on an i5-8250U CPU and inputs a set of color still life images to predict their scoring results. The average time for scoring a color still life work on this device is only 0.2s, and the error of the predicted score is within 5 points, which improves the scoring efficiency while ensuring accuracy. This demonstrates the effectiveness of the intelligent scoring mechanism proposed by the present invention.

[0100] like Figure 2 As shown in the figure, an intelligent painting scoring system based on attention convolutional network includes:

[0101] The dataset acquisition module is used to classify the collected paintings according to their levels to obtain the input dataset;

[0102] The network construction module builds a classification network based on global features and local features to obtain an attention convolutional network;

[0103] The framework construction module builds coarse and fine classifiers based on the attention convolutional network to obtain an intelligent scoring framework;

[0104] The assignment scoring module is used to input the input data set into the intelligent scoring framework to obtain the final score.

[0105] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0106] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for intelligently scoring paintings based on an attention convolutional network, characterized in that: The following steps are involved: S1), classify the collected paintings according to their levels to obtain an input data set; S2) Build a classification network based on global features and local features to obtain an attention convolutional network; specifically, it includes: Build an image embedding layer based on a two-dimensional convolutional neural network; Construct a local feature extraction layer based on depthwise separable convolutional blocks; Construct a global feature learning layer based on the Reshape function and multi-head attention mechanism module; Construct a fine-grained classification layer based on the global average pooling function and nonlinear activation function; The image embedding layer, local feature extraction layer, global feature learning layer and fine-grained classification layer are fitted to obtain the attention convolutional network; The depth-wise separable convolution block consists of two 1×1 point-by-point convolutions and one spatial depth convolution, and its formula is expressed as follows: Among them, Conv represents ordinary convolution, Y represents the output of the depth-separable convolution block, DWconv represents separable convolution, and B i Indicates the bias of each layer; The step of constructing a fine-grained classification layer based on the global average pooling function and the nonlinear activation function is expressed as follows: Among them, p j Represents the probability of the jth category, GHF represents the global average pooling output of local features and global features, represents the interactive information containing local features and global features, d represents the number of heads of the multi-head attention module, i represents the i-th depth-separable convolution block, n represents the number of depth-separable convolution blocks, and y j Represents the feature information of the jth category; S3) Building a coarse classifier and a fine classifier based on the attention convolutional network to obtain an intelligent scoring framework; S4) Input the input data set into the intelligent scoring framework to obtain the final score; specifically including: The image embedding layer based on the coarse classifier projects the input data set to obtain the multi-channel first-class feature space projection; Applying the local feature extraction layer of the coarse classifier to extract features from the first-category feature space projection of the multi-channel to obtain the first-category high-dimensional local features; The global feature learning layer of the coarse classifier is used to supplement the learning of the first type of high-dimensional local features to obtain the interactive information between the first type of local features and the first type of global features; Applying the fine-grained classification layer of the coarse classifier to classify the interactive information of the first-category local features and the first-category global features to obtain the first-category classification information; The image embedding layer based on the detailed classifier projects the first category classification information to obtain a multi-channel second category feature space projection; Applying the local feature extraction layer of the fine classifier to extract features from the second-category feature space projection of the multi-channel to obtain the second-category high-dimensional local features; The global feature learning layer of the detailed classifier is used to supplement the learning of the second type of high-dimensional local features to obtain the interactive information between the second type of local features and the second type of global features; Applying the fine-grained classification layer of the fine classifier to classify the interaction information of the second type of local features and the second type of global features to obtain the second type of classification information; The scoring module is applied to assign scores to the second category of classification information to obtain the final score.

2. The method for intelligently scoring paintings based on an attention convolutional network according to claim 1, characterized in that: Also includes: The coarse classifier and the fine classifier are trained respectively to obtain the trained coarse classifier and the fine classifier.

3. The method for intelligently scoring paintings based on an attention convolutional network according to claim 1, characterized in that: The step of constructing a coarse classifier and a fine classifier based on the attention convolutional network to obtain an intelligent scoring framework specifically includes: Based on the attention convolutional network, the depth-wise separable convolutional block is taken as the backbone network of the model to obtain a coarse classifier; Based on the attention convolutional network, the depth-wise separable convolutional block is used as the backbone network of the model to obtain a detailed classifier. Construct a scoring module based on random mood coefficient and assignment function; The coarse classifier, fine classifier and scoring module are fitted to obtain an intelligent scoring framework.

4. The method for intelligently scoring paintings based on an attention convolutional network according to claim 2, characterized in that: The step of training the coarse classifier and the fine classifier respectively to obtain the trained coarse classifier and the fine classifier specifically includes: Divide the input data set into a high- and low-class classification task to obtain a preheated data set; Pre-train the attention convolutional network model based on the preheated dataset to obtain the initial weights; The coarse classifier and the fine classifier are trained by gradient updating based on the initialized model parameters of the initial weights and the remaining input data set to obtain the trained coarse classifier and the fine classifier.

5. The method for intelligently scoring paintings based on an attention convolutional network according to claim 4, characterized in that: The step of training the coarse classifier and the fine classifier respectively to obtain the trained coarse classifier and the fine classifier further includes: Perform image enhancement on the input data set to obtain an enhanced input data set.

6. An intelligent painting scoring system based on attention convolutional network, characterized by: include: The dataset acquisition module is used to classify the collected paintings according to their levels to obtain the input dataset; The network construction module builds a classification network based on global features and local features to obtain an attention convolutional network; The framework construction module builds coarse and fine classifiers based on the attention convolutional network to obtain an intelligent scoring framework; The assignment scoring module is used to input the input data set into the intelligent scoring framework to obtain the final score; The network construction module constructs a classification network based on global features and local features to obtain an attention convolutional network; specifically, it includes: Build an image embedding layer based on a two-dimensional convolutional neural network; Construct a local feature extraction layer based on depthwise separable convolutional blocks; Construct a global feature learning layer based on the Reshape function and multi-head attention mechanism module; Construct a fine-grained classification layer based on the global average pooling function and nonlinear activation function; The image embedding layer, local feature extraction layer, global feature learning layer and fine-grained classification layer are fitted to obtain the attention convolutional network; The depth-wise separable convolution block consists of two 1×1 point-by-point convolutions and one spatial depth convolution, and its formula is expressed as follows: Among them, Conv represents ordinary convolution, Y represents the output of the depth-separable convolution block, DWconv represents separable convolution, and B i Indicates the bias of each layer; The step of constructing a fine-grained classification layer based on the global average pooling function and the nonlinear activation function is expressed as follows: Among them, p j Represents the probability of the jth category, GHF represents the global average pooling output of local features and global features, represents the interactive information containing local features and global features, d represents the number of heads of the multi-head attention module, i represents the i-th depth-separable convolution block, n represents the number of depth-separable convolution blocks, and y j Represents the feature information of the jth category; The assignment scoring module inputs the input data set into the intelligent scoring framework to obtain the final score; specifically, it includes: The image embedding layer based on the coarse classifier projects the input data set to obtain the multi-channel first-class feature space projection; Applying the local feature extraction layer of the coarse classifier to extract features from the first-category feature space projection of the multi-channel to obtain the first-category high-dimensional local features; The global feature learning layer of the coarse classifier is used to supplement the learning of the first type of high-dimensional local features to obtain the interactive information between the first type of local features and the first type of global features; Applying the fine-grained classification layer of the coarse classifier to classify the interactive information of the first-category local features and the first-category global features to obtain the first-category classification information; The image embedding layer based on the detailed classifier projects the first category classification information to obtain a multi-channel second category feature space projection; Applying the local feature extraction layer of the fine classifier to extract features from the second-category feature space projection of the multi-channel to obtain the second-category high-dimensional local features; The global feature learning layer of the detailed classifier is used to supplement the learning of the second type of high-dimensional local features to obtain the interactive information between the second type of local features and the second type of global features; Applying the fine-grained classification layer of the fine classifier to classify the interaction information of the second type of local features and the second type of global features to obtain the second type of classification information; The scoring module is applied to assign scores to the second category of classification information to obtain the final score.

Images