A deep learning-based sketch work intelligent evaluation method

Abstract

The application provides a sketch work intelligent evaluation method based on deep learning, and belongs to the technical field of image evaluation. The method comprises the following steps: acquiring a data set; constructing an evaluation system composed of a modeling evaluation module, a sketch relationship evaluation module and an artistic expression evaluation module; training an optimal evaluation system and deploying the optimal evaluation system on a server side; uploading art examination type sketch works and corresponding topic pictures to the server side, and uploading the art examination type sketch works meeting the evaluation requirements and the corresponding topic pictures to the optimal evaluation system; evaluating the art examination type sketch works meeting the evaluation requirements and the corresponding topic pictures, and returning the evaluation results to a user side. The application solves the problems of unbalanced resource allocation, high cost and slow feedback in the traditional art examination drawing training scoring mode, and solves the problems of poor pertinence and low precision of the existing image aesthetic quality evaluation method in this application environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image evaluation, and particularly relates to a sketch work intelligent evaluation method based on deep learning. BACKGROUND

[0002] With the continuous development of artificial intelligence technology (AI), AI is widely applied in many fields and has a profound impact on the development of various fields. Among them, AI performs outstandingly in the field of education, and the involvement of AI has greatly enriched the education approach and optimized resource allocation. Therefore, it is of great social significance and practical value to explore the application of artificial intelligence technology represented by deep learning in the auxiliary teaching of art examination drawing under this perspective.

[0003] In the drawing art examination preparation process, similar to cultural examinees, art examinees still need a lot of "mechanized" daily training and practice. However, the traditional "one-to-one" or "one-to-many" work scoring mode relying on the review of agency teachers cannot meet the needs of art drawing examinees for "repeated multiple times" and "timely feedback", and the cost is also high. Moreover, the characteristics of the application scenario are that the subjective weight of artistic evaluation is weakened and the objective proportion is greatly improved, which provides an opportunity for the involvement and role of artificial intelligence. SUMMARY

[0004] In view of the above deficiencies in the prior art, the application provides an intelligent sketch work evaluation method based on deep learning, which solves the problems of unbalanced resource allocation, high cost, and slow feedback in the traditional art examination drawing training scoring mode, and solves the problems of poor pertinence and low precision of existing image aesthetic quality evaluation methods in this application environment, thereby effectively improving the learning efficiency and learning experience of art examination drawing students and exploring the application of artificial intelligence in the field of art education.

[0005] In order to achieve the above purposes, the technical scheme adopted by the application is:

[0006] The present application provides an intelligent sketch work evaluation method based on deep learning, comprising the following steps:

[0007] S1, collect a plurality of art examination sketch works, and screen and label them to obtain a data set;

[0008] S2, construct an evaluation system composed of a modeling evaluation module, a sketch relationship evaluation module, and an artistic expression evaluation module;

[0009] S3, train the evaluation system using the data set to obtain an optimal evaluation system, and deploy it on the server side;

[0010] S4, upload the sketch works for art examination and the subject pictures corresponding to the sketch works to a server end, and upload the sketch works for art examination meeting the evaluation requirements and the subject pictures corresponding to the sketch works to an optimal evaluation system;

[0011] S5, evaluate the sketch works for art examination meeting the evaluation requirements and the subject pictures corresponding to the sketch works by using the optimal evaluation system, and return the evaluation results to the user end, and complete the intelligent evaluation of the sketch works.

[0012] The application has the following advantages compared with the traditional correction: 1. It is more convenient and real-time, and can be evaluated at any time and any place as long as there is a mobile phone and the Internet, and the progress and shortcomings can be found in time. 2. The cost is lower. The application can correct the sketch works for art examination at a low cost. Compared with directly analyzing the sketch works for art examination by using an image aesthetic quality model, the application is specially developed for the sketch works for art examination, has higher accuracy and stronger generality, and adopts CNN and DNN based on the attention mechanism, so that the advantages of the two are combined, and the sketch works for art examination are evaluated from the global and the details.

[0013] Further, the step S1 comprises the following steps:

[0014] S101, collect a plurality of sketch works for art examination, and screen the sketch works;

[0015] S102, evaluate in sequence from the sketch relationship, the modeling and the artistic expressiveness, and obtain a data set.

[0016] The above further scheme has the following advantages: the application effectively obtains a rich and high-quality data set by screening and marking the works.

[0017] Further, the modeling evaluation module in the step S2 comprises a face key point extraction sub-module, a semantic segmentation object position key point extraction sub-module and a sketch modeling evaluation sub-module.

[0018] The face key point extraction sub-module is used for extracting the face key points in the sketch works for art examination of the portrait type.

[0019] The semantic segmentation object position key point extraction sub-module is used for extracting the position key point coordinates of each object in the subject pictures and the sketch works in all the sketch works for art examination except the portrait type, and storing the position key point coordinates in the same object order, wherein the key points of the sketch works are stored first and then the key points of the subject pictures are stored.

[0020] The sketch modeling evaluation submodule is configured to extract key position coordinates of the subject picture and the sketch work through the face key point or the object position key point, and obtain an evaluation grade of the sketch work modeling dimension.

[0021] The above further scheme has the beneficial effect that the key point method is adopted, and the essential factors affecting the sketch modeling can be captured, and the generalization ability is stronger.

[0022] Further, the loss function of the face key point extraction submodule is:

[0023]

[0024]

[0025] wherein, represents the loss function of the face key point extraction part, represents a predicted value of the predicted face key point coordinates, represents a true value of the face key point coordinates, all are multi-dimensional vectors, represents the loss value corresponding to the element in the i-th row and the j-th column, represents the index of the element, represents the predicted value corresponding to the element in the i-th row and the j-th column, represents the true value corresponding to the element in the i-th row and the j-th column; represents the total number of elements in the matrix, represents a threshold parameter of the smooth L1 norm loss; The activation function of the face key point extraction submodule is:

[0026]

[0027]

[0028]

[0029] wherein, represents the activation function of the face key point extraction part, represents an input value, represents a Gaussian error function, represents a natural exponential, represents a differential of h, represents an integral variable.

[0030] ​​​​Further, the sketch relationship evaluation module and the artistic expression evaluation module in the step S2 both use a CNN model and a DNN model based on a transformer architecture to extract features.

[0031] The CNN model extracts features by using a convolution operation.

[0032] The DNN model based on the TRANSFORMER architecture extracts features by using a multi-head attention mechanism, and combines the output results of the CNN model and the DNN model by using a weight coefficient, wherein the expression of the output results is as follows:

[0033]

[0034] Among them, represents the output results of the DNN model based on the TRANSFORMER architecture and the CNN model, represents the weight coefficient, represents the result of the CNN model, represents the result of the DNN model based on the TRANSFORMER architecture.

[0035] The above further scheme has the beneficial effects that: for the evaluation of the sketch for art examination, the present application not only focuses on the details, but also must be grasped from the whole, and the output results of the two are combined by using a weight coefficient, and the specific formula is: the advantages of the two are combined, and good results can be achieved.

[0036] Further, the optimizer used by the modeling evaluation module, the sketch relationship evaluation module and the artistic expression evaluation module is:

[0037]

[0038]

[0039]

[0040]

[0041] Among them, represents the first-order moment estimation, represents the updated , represents the exponential decay rate of the first-order moment estimation, and is 0.898, represents the gradient of the parameter, represents the second-order moment estimation, represents the updated , represents the exponential decay rate of the second-order moment estimation, and is 0.998, learning rate, parameters that the optimizer needs to optimize, updated , value that maintains numerical stability, taken as 1e-08, , of iteration, of iteration, iteration, iteration number, initially 0, and increases by 1 each time.

[0042] Further, the step S5 comprises the following steps:

[0043] S501, gradient cropping the art examination type sketch works meeting the evaluation requirements and the corresponding topic pictures to obtain nine part pictures;

[0044] S502, evaluating the complete picture by using the optimal evaluation system to obtain three-dimensional evaluation, and evaluating the nine part pictures by using the sketch relationship evaluation module and artistic expressiveness evaluation module in the optimal evaluation system to obtain two-dimensional evaluation;

[0045] S503, generating corresponding comments according to the corresponding dimensional grades of the complete picture;

[0046] S504, returning the three-dimensional grades of the complete picture and the two-dimensional grades of the nine part pictures, and the generated comments to the user end to complete the intelligent evaluation of the sketch works.

[0047] Further, the gradient cropping is:

[0048]

[0049] wherein, output, input multi-dimensional gradient tensor, L2 norm of , L2 norm of cropping threshold, the th element in , index number of the element, total number of elements in , predicted value of the predicted face key point coordinate.

[0050] The beneficial effect of the further scheme is that gradient clipping can prevent gradient explosion problem and prevent the model from passing the optimal point. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of the method of the present application.

[0052] Figure 2 A flowchart of the processing of the sketch works for art examination in the present embodiment.

[0053] Figure 3 A flowchart of the processing of the modeling evaluation module in the present application.

[0054] Figure 4 A flowchart of the processing of the sketch relationship evaluation module or the artistic expressiveness evaluation module in the present application. DETAILED DESCRIPTION

[0055] The specific embodiments of the present application are described below to facilitate the understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0056] EMBODIMENT

[0057] In the present embodiment, the present application collects a large number of sketch works for art examination, performs screening and labeling to obtain a high-quality labeled data set, adopts the idea of artificial intelligence, utilizes the characteristics of deep learning that can effectively identify image features, and constructs a core evaluation system composed of a modeling evaluation module, a line, block, volume, dynamic modeling art expression evaluation module (hereinafter referred to as: sketch relationship evaluation module), and an object light and shadow, light and dark, black and white level and picture black and white rhythm, and the overall feeling and artistic feeling evaluation module of the whole picture (hereinafter referred to as: artistic expressiveness evaluation module). The obtained data set is repeatedly trained to obtain an evaluation system meeting the predetermined accuracy, which is deployed on the server side. The user uploads the sketch works for art examination and the corresponding topic pictures to the server side through the user side, the server receives the pictures, first judges whether the sketch works pictures meet the evaluation requirements, if they meet, sends them to the cutting module after sending them to the core evaluation system, evaluates the works from three aspects, returns the results to the user side, and at the same time, according to different results, there will be corresponding comments to help the user progress, such as Figures 1-2 As shown in the figure, the present application provides a sketch work intelligent evaluation method based on deep learning, and the implementation method is as follows:

[0058] S1, collect a number of art exam sketch works, and screen and label them to obtain data, the implementation method is as follows:

[0059] S101, collect a number of art exam sketch works, and screen them;

[0060] S102, evaluate in turn in three aspects of sketch relationship, modeling and artistic expression to obtain a data set.

[0061] In this embodiment, after collecting a large number of art exam sketch works and screening them, the evaluation and labeling are performed according to the requirements of art exam sketch (evaluation from three aspects, in turn, sketch relationship, modeling, and artistic expression), each evaluation dimension is divided into four grades A, B, C, and D, corresponding to excellent, good, qualified, and unqualified in art exam sketch evaluation. After completing the screening and labeling work, a rich and high-quality data set is obtained.

[0062] S2, construct an evaluation system composed of a modeling evaluation module, a sketch relationship evaluation module and an artistic expression evaluation module;

[0063] In this embodiment, the evaluation system includes a modeling evaluation module, a sketch relationship evaluation module and an artistic expression evaluation module. The input is a sketch picture and a title picture, and the output is a grade (A, B, C, D) in three dimensions.

[0064] The modeling evaluation module includes a face key point extraction submodule, a semantic segmentation object position key point extraction submodule and a sketch modeling evaluation submodule;

[0065] The face key point extraction submodule is used to extract the face key points in the art exam portrait sketch;

[0066] The loss function of the face key point extraction submodule is:

[0067]

[0068]

[0069] wherein, represents the loss function of the face key point extraction part, represents the predicted value of the predicted face key point coordinates, represents the true value of the face key point coordinates, all are multi-dimensional vectors, represents the loss value corresponding to the element in the index of the item, represents the loss value corresponding to the a predicted value corresponding to the item element, denotes the total number of items in the a true value corresponding to the item element; denotes the total number of items in the denotes a threshold parameter of the smooth L1 norm loss;

[0070] The activation function of the face key point extraction submodule is:

[0071]

[0072]

[0073] wherein, denotes the activation function of the face key point extraction part, denotes the input value, denotes the Gaussian error function, denotes the natural exponential, denotes the differential of h, denotes the integral variable;

[0074] The semantic segmentation object position key point extraction submodule is used to extract the position key point coordinates of each object in the subject picture and the sketch work in all the art examination sketch categories except the human head portrait category, and store them in the same object order, wherein the sketch work key points are stored first and then the subject picture key points are stored in the storage process.

[0075] The sketch modeling evaluation submodule is used to extract the key position coordinates of the subject picture and the sketch work through the face key points or the object position key points, so as to obtain the evaluation grade of the sketch work modeling dimension.

[0076] The modeling evaluation module: this module is composed of the face key point submodule, the semantic segmentation object position key point extraction submodule, and the sketch modeling evaluation submodule according to the key point coordinates.

[0077] The advantages of adopting this processing flow in this embodiment compared to directly using a deep learning classifier to analyze the sketch and the theme picture to obtain the result are: 1) greatly reducing the amount of calculation, using the foregoing method, the sketch and the theme picture need to be spliced and sent into the classifier, while the present method only needs to extract the key point coordinates of each picture separately, and the coordinates are analyzed to obtain the grade, and the amount of coordinate data is greatly reduced compared to the amount of picture data. 2) effectively improving the accuracy. The existing face key point or semantic segmentation dataset has a large amount of data and high quality, and the accuracy of the deep learning model increases with the size of the dataset. The size of the existing art examination sketch dataset is difficult to match. 3. Better generalization ability. Directly using a deep learning classifier to analyze the sketch and the theme picture to obtain the result will cause a large amount of irrelevant information to be included in the factors that determine the grade, causing overfitting, and only being effective for art examination sketches similar to the training set. The present method uses key points, which can better capture the essential factors affecting the sketch modeling, and has better generalization ability.

[0078] In this embodiment, the face key point extraction submodule: this part is suitable for art examination portrait sketch (including half-length portrait) (this type of sketch accounts for a large proportion in art examination sketch examination, so it is developed separately according to its characteristics), which is composed of a DNN model based on TRANSFORMER architecture based on attention mechanism, and the loss function used is the SmoothL1 loss function.

[0079] In this embodiment, the face key point extraction submodule can identify 106 face key points (including eyes, ears, nose, jaw, mouth, etc.) with high accuracy, completely covering the modeling evaluation feature points of the portrait class of art examination sketches, and extracting 106 face key points from the sketch picture and the theme picture respectively, and storing them in the same order. The sketch key points are stored first, and then the theme picture key points are stored. To improve the accuracy, this part is first trained using an authoritative public face key point dataset, and then fine-tuned using part of the pictures in the dataset in (1) (which have been labeled).

[0080] In this embodiment, the object position key point extraction submodule is suitable for all art examination sketch categories except the portrait class. The semantic segmentation model (such as Swin-transformer) is used to extract the position key point coordinates of each object in the theme picture and the sketch (a series of points for the boundary of the object), and store them in the same object order. The sketch key points are stored first, and then the theme picture key points are stored. To improve the accuracy, this part is also first trained using a high-quality public dataset containing a large number of common objects, and then fine-tuned using part of the pictures in the dataset in (1) (which have been labeled). The loss function and the activation function can be used for semantic segmentation tasks.

[0081] In this embodiment, the sketch modeling evaluation submodule is performed according to the key point coordinates. This part uses a deep learning-based table data feature extraction model (such as FT-Transformer, Resnet). Compared with traditional table data extraction models, the deep learning-based model is more suitable for the present application due to its stronger learning and generalization capabilities, especially for the characteristics of the objects drawn in the sketches of other categories except for the head portrait type of the art examination sketch, which vary greatly in quantity, size, category, and form, making it difficult to learn their essential features. Through repeated training of the data set obtained by (1) on this module until the expected accuracy is obtained, this module can extract the key position coordinates of the sketch work and the topic picture through the face key points (or object position key points), and obtain the evaluation level of the sketch work modeling dimension.

[0082] In this embodiment, as shown in Figures 3-4 , the technical implementation of the sketch relationship evaluation module and the artistic expression evaluation module is the same, which is different in evaluation dimension, so they are separated. Here, the technical implementation is described uniformly. The two modules are implemented by a CNN model with a large number of parameters (the parameter quantity is at least 100M) and a DNN model based on TRANSFORMER architecture based on attention mechanism. The CNN model is good at capturing local details, and the DNN model based on TRANSFORMER architecture based on attention mechanism is good at extracting global information. For the evaluation of art examination sketches, both the details and the overall situation must be grasped. By combining the output results of the two through a weight coefficient, the specific formula is as follows , the advantages of the two can be combined to achieve good results, where the loss functions of the CNN and the DNN model based on TRANSFORMER architecture are cross-entropy loss functions :

[0083]

[0084] wherein, represents the output results of the DNN model based on TRANSFORMER architecture and the CNN model, represents the weight coefficient, represents the result of the CNN model, represents the result of the DNN model based on TRANSFORMER architecture.

[0085]

[0086] wherein, represents the cross-entropy loss function, represents the total number of samples, represents the sample index number, Indicates the total number of categories. Indicates the category index number. Indicates sample Is it If it is a class, take 1; otherwise, take 0. This indicates taking the natural logarithm. Indicates sample for The probability of.

[0087] In this embodiment, the sketching relationship evaluation module and the artistic expression evaluation module are trained using a dataset. The deep learning models corresponding to the three modules of the core evaluation system all use the Adam optimizer, with an initial learning rate of 0.001, beta1 of 0.898, and beta2 of 0.998.

[0088] The optimizer used in the form evaluation module, the sketch relationship evaluation module, and the artistic expression evaluation module is:

[0089]

[0090]

[0091]

[0092]

[0093] in, This represents the first-order moment estimate. Indicates the updated , Let represent the exponential decay rate of the first-moment estimate, taken as 0.898. Represents the gradient of the parameters. This indicates the second-order moment estimation. Indicates the updated , Let represent the exponential decay rate of the second-order moment estimate, taken as 0.998. Indicates the learning rate. This indicates the parameters that the optimizer needs to optimize. Indicates the updated , The value representing the preservation of numerical stability is taken as 1e-08. express of iteration express of iteration This represents the number of iterations, initially set to 0, and incremented by 1 with each iteration.

[0094] S3, train the evaluation system using the data set to obtain an optimal evaluation system and deploy the optimal evaluation system on a server side;

[0095] S4, upload the sketch works for art examination and the subject pictures corresponding to the sketch works to the server side, and upload the sketch works for art examination meeting the evaluation requirements and the subject pictures corresponding to the sketch works to the optimal evaluation system;

[0096] In the embodiment, the picture meeting the requirements of the sketch work image quality evaluation is input into a sketch work force image quality evaluation module, which is a relatively simple CNN model (selectable models include ResNet50, EfficientNet, etc.), the input is the sketch work picture, and the output is 1 (meeting the evaluation requirements) or 0 (not meeting the evaluation requirements). The normal sample is the data set created in (I), and the abnormal sample is an abnormal picture such as blur, tilt, overexposure, etc.

[0097] S5, evaluate the sketch works for art examination meeting the evaluation requirements and the subject pictures corresponding to the sketch works using the optimal evaluation system, and return the evaluation results to the user side to complete the intelligent evaluation of the sketch works, and the implementation method is as follows:

[0098] S501, gradient cropping is performed on the sketch works for art examination meeting the evaluation requirements and the subject pictures corresponding to the sketch works to obtain nine partial pictures;

[0099] S502, evaluate the complete picture using the optimal evaluation system to obtain three-dimensional evaluation, and evaluate the nine partial pictures using the sketch relationship evaluation module and the artistic expression evaluation module in the optimal evaluation system to obtain two-dimensional evaluation;

[0100] S503, generate corresponding comments according to the corresponding dimensional grades of the complete picture;

[0101] S504, return the three-dimensional grades of the complete picture and the two-dimensional grades of the nine partial pictures and the generated comments to the user side to complete the intelligent evaluation of the sketch works.

[0102] In the embodiment, the sketch work picture and the subject picture are cut into nine small parts in a nine-square grid manner by the cutting module, and the nine small parts are sent to the sketch relationship evaluation module and the artistic expression evaluation module in the core evaluation system for evaluation. The complete picture can be sent to the core evaluation system for evaluation. The nine small parts can help the user quickly locate the problems in the sketch relationship and artistic expression, but cannot perform modeling evaluation on the nine small parts; the accuracy of the evaluation scores of the three aspects of the complete picture will be higher.

[0103] Gradient cropping is:

[0104]

[0105] wherein, represents an output, represents an input multi-dimensional gradient tensor, represents an L2 norm of an L2 norm of a clipping threshold, represents an i-th item in an i-th item in an index number of an item.

[0106] In this embodiment, the comment is a general guidance given for different categories, different dimensions, and different grades, and is output according to the category, dimension, and grade of the evaluated sketch.

[0107] In this embodiment, as shown in Figure 2 , the user uploads the sketch to be evaluated and the corresponding question picture to the server end through the user end, the server end determines whether the picture quality meets the requirements after receiving the picture, if it meets the requirements, it is sent to the cutting module, the sketch and the question picture are cut into nine part pictures, the complete picture and the nine part pictures are sent to the core evaluation system for evaluation, the three-dimensional grades of the complete picture and the nine part pictures are obtained, the grade corresponding to the complete picture is sent to the comment generation system to generate a comment, finally, the three-dimensional grades of the complete picture and the nine part pictures and the comment are transmitted back to the client end.

Claims

1. A deep learning-based intelligent evaluation method for sketches, characterized in that, Includes the following steps: S1. Collect a number of art exam-related sketches, filter and annotate them to obtain a dataset; S2. Construct an evaluation system consisting of a modeling evaluation module, a sketching relationship evaluation module, and an artistic expression evaluation module; The model evaluation module in step S2 includes a facial key point extraction submodule, a semantic segmentation object location key point extraction submodule, and a sketch model evaluation submodule. The facial key point extraction submodule is used to extract facial key points in art exam portrait sketches. The semantic segmentation object location key point extraction submodule is used to extract the location key point coordinates of each object in the question image and the sketch in all art exam sketch categories except for human head portraits, and store them in the same object order. During the storage process, the key points of the sketch are stored first and then the key points of the question image are stored. The sketch modeling evaluation submodule is used to extract the key position coordinates of the title image and the sketch work by using key points of the face or key points of the object position, and to obtain the evaluation level of the modeling dimension of the sketch work. In step S2, both the sketch relationship evaluation module and the artistic expression evaluation module use CNN models and DNN models based on transformer architecture to extract features. The CNN model uses convolution operations to extract features; The DNN model based on the TRANSFORMER architecture extracts features using a multi-head attention mechanism and combines the outputs of the CNN model and the DNN model through a weight coefficient. The expression for the output is as follows: in, This represents the output results of DNN and CNN models based on the TRANSFORMER architecture. Indicates the weighting coefficient. This represents the result of the CNN model. This represents the results of a DNN model based on the TRANSFORMER architecture; S3. Train the evaluation system using the dataset to obtain the optimal evaluation system, and deploy it on the server. S4. Upload the art exam sketches and their corresponding question images to the server, and upload the art exam sketches and their corresponding question images that meet the evaluation requirements to the best evaluation system. S5. Use the optimal evaluation system to evaluate the art exam sketches that meet the evaluation requirements and their corresponding question images, and return the evaluation results to the user to complete the intelligent evaluation of the sketches. Step S5 includes the following steps: S501. Perform gradient cropping on the art exam sketches that meet the evaluation requirements and their corresponding question images to obtain images in nine parts. S502. The complete image is evaluated using the optimal evaluation system to obtain a three-dimensional evaluation. The nine parts of the image are evaluated using the sketch relationship evaluation module and the artistic expression evaluation module of the optimal evaluation system to obtain a two-dimensional evaluation. S503. Generate corresponding comments based on the corresponding dimension level of the complete image; S504. Return the three-dimensional rating of the complete image and the two-dimensional rating of the nine partial images, along with the generated comments, to the user to complete the intelligent evaluation of the sketch.

2. The intelligent evaluation method for sketches based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps: S101. Collect a number of art exam-related sketches and screen them. S102. The dataset is obtained by evaluating the three aspects of sketching relationships, form, and artistic expression.

3. The intelligent evaluation method for sketches based on deep learning according to claim 1, characterized in that, The loss function of the facial landmark extraction submodule is: in, This represents the loss function for facial landmark extraction. This represents the predicted value of the facial landmark coordinates. This represents the true values ​​of the facial landmark coordinates. Both are multidimensional vectors. express The Middle The loss value corresponding to each item element, Indicates the index of the item. express The Middle The predicted value corresponding to the item element, express The Middle The actual value corresponding to the item element; express The total number of middle terms, The threshold parameter representing the smoothing L1 norm loss; The activation function of the facial landmark extraction submodule is: in, This represents the activation function used for extracting facial landmarks. Indicates the input value. Represents the Gaussian error function. Represents the natural index. Denotes the differential with respect to h. This represents the integral variable.

4. The intelligent evaluation method for sketches based on deep learning according to claim 1, characterized in that, The optimizer used in the form evaluation module, the sketch relationship evaluation module, and the artistic expression evaluation module is: in, This represents the first-order moment estimate. Indicates the updated , Let represent the exponential decay rate of the first-moment estimate, taken as 0.

898. Represents the gradient of the parameters. This indicates the second-order moment estimation. Indicates the updated , Let represent the exponential decay rate of the second-order moment estimate, taken as 0.

998. Indicates the learning rate. This indicates the parameters that the optimizer needs to optimize. Indicates the updated , The value representing the preservation of numerical stability is taken as 1e-08. express of iteration express of iteration This represents the number of iterations, initially set to 0, and incremented by 1 with each iteration.

5. The intelligent evaluation method for sketches based on deep learning according to claim 1, characterized in that, The gradient clipping is as follows: in, Indicates the output. This represents the input multidimensional gradient tensor. express L2 norm, express The L2 norm needs to be pruned to a threshold. express The first in item, Indicates the index number of the item. express The total number of middle terms, This represents the predicted value of the coordinates of facial landmarks.

Images