Drawing image scoring method and system based on deep learning

Abstract

The invention provides a deep learning-based drawing image scoring method and system. The deep learning-based drawing image scoring method comprises the steps of S1, collecting drawing images and image label data as training samples; s2, preprocessing the training sample data, training an output model structure of a neural convolutional network, and identifying visual description and feature description of a drawing graph; s3, preprocessing a drawing image file of a student, and then converting the image into visual description of an image level and an object level through a training model; and S4, generating a scoring instruction according to the scoring rule, and using a specific instruction to guide the large language model to evaluate the image according to a specific standard. According to the drawing image automatic scoring system provided by the invention, the scoring efficiency and accuracy are improved through an automatic and intelligent evaluation method, meanwhile, the number of constituent elements inconsistent with text prompts in the image is identified and counted, and detailed scores and reasons are provided.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a drawing image scoring method and system based on deep learning. Background Art

[0002] With the advancement of computer vision and machine learning technologies, the ability of computers to automatically process and analyze images has been significantly improved.

[0003] Traditional image scoring methods are mostly manual evaluations, which are inefficient and highly subjective. Traditional automated image quality evaluation methods aim to use computers to simulate the human visual system to automatically evaluate the degree of image distortion. Although they can obtain the same evaluation value as the subjective evaluation results, they are insufficient in simulating human understanding of image content and object-level alignment between images and evaluation rules. In order to improve the automation and objectivity of image scoring, a software system based on image processing and analysis is needed that can automatically score and provide accurate evaluation results. Summary of the invention

[0004] One purpose of the present invention is to propose a method and system for scoring drawing images based on deep learning. The automatic painting scoring system based on deep learning proposed by the present invention solves many defects of the prior art and improves the efficiency, accuracy and interpretability of scoring through automated and intelligent methods.

[0005] A drawing image scoring system based on deep learning according to an embodiment of the present invention includes:

[0006] The preprocessing module is used to perform preliminary processing on the input image, including grayscale, image enhancement, denoising, normalization, and feature extraction to ensure the accuracy of subsequent analysis;

[0007] Model training is used to perform deep learning training on various painting images and image description label data to obtain an evaluation model for image content description, which is mainly responsible for converting images into multi-level, object-centric visual descriptions;

[0008] The scoring rule conversion module is used to convert the scoring rules input by the user into a series of evaluation instructions;

[0009] The scoring module uses large language models to measure the alignment between the generated image and text based on the obtained image visual description, and uses their reasoning ability to understand the combined properties of objects and the complex interactions between multiple objects. Based on the above generated visual description, the entire evaluation process consists of three steps: instruction follow-up scoring, rule-enhanced scoring, and scoring reason generation;

[0010] The output module performs weighted calculations on the scores of each instruction based on the scoring instructions, and finally outputs the scores of the students' paintings and the reasons for the scores.

[0011] A drawing image scoring method based on deep learning according to an embodiment of the present invention comprises the following steps:

[0012] S1, collect painting images and image content description label training samples;

[0013] S2. Perform grayscale conversion, image enhancement, noise reduction, normalization, and image feature extraction on the training paintings, and train the model structure for generating image visual description and feature description;

[0014] S3, using the model of image visual description and feature description to evaluate the painting to be evaluated, obtaining image visual description information and content feature information, and using the language macro model to merge the description information and feature information into an object-centered visual description;

[0015] S4. Use the LLM model to convert the scoring rules into scoring instructions. By providing the LLM model with scoring instructions and image visual description information, the LLM model evaluates the image according to the scoring instructions to obtain the final score of the painting and the reason for the scoring.

[0016] Optionally, the S1 includes:

[0017] S11, collecting training sample painting images;

[0018] S12. Classify and annotate the painting images: image attributes, content description, relationship between content objects, aesthetic evaluation, and emotional evaluation.

[0019] Optionally, S2 includes:

[0020] S21, grayscale the training sample image, and convert the RGB value of each pixel of the image into a grayscale value of R=G=B:

[0021] Gray(i,j)=0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j);

[0022] Where i, j represent the coordinate position of the pixel, Gray(i, j) is the gray value after processing, R(i, j), G(i, j), and B(i, j) represent the RGB values ​​of the pixel respectively;

[0023] S22, performing image enhancement on the training sample image, changing the grayscale distribution of image pixels, enhancing global or local useful information in the image, improving image quality, and enhancing image clarity;

[0024] S23, performing noise reduction processing on the training sample image;

[0025] S24, normalizing the training sample image, changing the size of the image, and changing the image pixel size to a fixed size, the size must be consistent with the requirements of the convolutional neural network;

[0026] S25, extracting features of training sample images;

[0027] S26. Model training. The deep learning framework uses the transform framework and the pre-trained BLIP model with 2.7B parameters to perform multiple rounds of training on the image data after data preprocessing and the image label data to obtain a network model that can quickly return the image visual description information and content feature information. The network model is then saved for subsequent painting scoring testing and operation.

[0028] Optionally, the S23 includes:

[0029] S231, grouping similar blocks, selecting some reference blocks in the noise image, and searching for similar blocks within a certain range near the reference blocks, calculating the similarity between these blocks and the reference blocks, and grouping them according to the similarity:

[0030]

[0031] Where A= ( a1,a2,,a n) and B = ( b1,b2,,b n) are two points in n-dimensional space, d ( A,B ) The Euclidean distance between point A and point B represents the similarity between patches;

[0032] S232, for each grouped three-dimensional matrix, perform two-dimensional transformation using DCT transformation, then perform one-dimensional transformation in the third dimension using Hadamard transformation, perform hard threshold processing on the transformed matrix, set coefficients less than the threshold to 0, and finally obtain the processed image block through inverse transformation;

[0033] S233, combining the processed image blocks into the original image, performing weighted average of the intensity of each pixel by the value of the block at the corresponding position, and obtaining a basic estimated image through the above steps.

[0034] Optionally, the S25 includes:

[0035] S251, acquiring an image and identifying pixel information of an input image, for example, the pixels of the input image are 100*200 pixels;

[0036] S252, image segmentation, image token calculation: the image is divided into multiple 20*20 pixel blocks, the height of the image is divided into 100 / 20=5 row blocks, the width of the image is divided into 200 / 20=10 column blocks, so the total number of blocks is 5*10=50 blocks, each block is flattened into a vector of length 20*20=400, so we have 50 image tokens of length 400;

[0037] S253, linear embedding block, maps each image token of length 400 to an embedding space through a linear transformation, generating an embedding vector of length 768. 768 is the dimension or length of each token after tokenization. After being flattened and linearly transformed, each image block will be mapped to an embedding vector of length 768, which means that the representation of each token is in a 768-dimensional embedding space, so each token is represented as an embedding vector of length 768.

[0038] S254, initialize a token with all 0s as the prediction token;

[0039] S255, concatenate the predicted token with 50 image tokens to form a token sequence. The sequence is in the form of: [ Prediction Token, Image Token1, Image Token2, ..., Image Token50 ] ;

[0040] S256: Add position embedding to all tokens. Position embedding is a vector with the same length as the token sequence, which is used to indicate the position of each token in the sequence.

[0041] S257, image block encoding: normalize the image blocks to ensure that the input data is evenly distributed in different dimensions, use a multi-head self-attention mechanism to capture the relationship between tokens, and obtain the global features of the image. Normalize the data after self-attention processing to ensure that the processed data remains evenly distributed. Finally, use a feedforward neural network: consisting of a fully connected layer, an activation layer, and another fully connected layer to convert the token information. Each token is fully processed and converted to capture the rich feature information of the input image.

[0042] Optionally, the S3 includes:

[0043] S31, preprocessing the painting image, and obtaining a global description, a local description, and a feature description of the image based on the training model generated in S2;

[0044] S32. Use a large language model to merge the global description, local description and feature description of the image into an object-centric visual description.

[0045] Optionally, the S4 includes:

[0046] S41. Through the LLM model, interpret the scoring rules and generate instructions that can be used for scoring. Subsequent scoring uses the scoring instructions.

[0047] S42: Based on the visual description information centered on the image object obtained in S3 and the generated scoring instructions, the powerful reasoning ability of the LLM model is used to perform scoring reasoning one by one according to the instructions to obtain the score and scoring details.

[0048] The beneficial effects of the present invention are:

[0049] The deep learning-based automatic painting scoring system proposed in this invention can significantly improve the efficiency and consistency of image scoring. Through automated image processing and analysis, it follows the given scoring rules to perform multi-dimensional and multi-granular combined scoring, and provides well-reasoned scores, while reducing the subjectivity of manual evaluation, and can also process a large amount of image data, thus improving the accuracy and reliability of the evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0051] Figure 1 A drawing image scoring method based on deep learning and a system architecture diagram in the system proposed by the present invention;

[0052] Figure 2 This is a flowchart of the training of an image visual description information extraction model in a deep learning-based drawing image scoring method proposed in the present invention. DETAILED DESCRIPTION

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0054] refer to Figure 1 and Figure 2 , a drawing image scoring method and system based on deep learning, comprising:

[0055] The preprocessing module is used to perform preliminary processing on the input image, including grayscale, image enhancement, denoising, normalization, and feature extraction to ensure the accuracy of subsequent analysis;

[0056] Model training is used to perform deep learning training on various painting images and image description label data to obtain an evaluation model for image content description, which is mainly responsible for converting images into multi-level, object-centric visual descriptions;

[0057] The scoring rule conversion module is used to convert the scoring rules input by the user into a series of evaluation instructions;

[0058] The scoring module uses large language models to measure the alignment between the generated image and text based on the obtained image visual description, and uses their reasoning ability to understand the combined properties of objects and the complex interactions between multiple objects. Based on the above generated visual description, the entire evaluation process consists of three steps: instruction follow-up scoring, rule-enhanced scoring, and scoring reason generation;

[0059] The output module performs weighted calculations on the scores of each instruction based on the scoring instructions, and finally outputs the scores of the students' paintings and the reasons for the scores.

[0060] In this embodiment, a drawing image scoring method based on deep learning includes:

[0061] S1, collect painting images and image content description label training samples;

[0062] S2. Perform grayscale conversion, image enhancement, noise reduction, normalization, and image feature extraction on the training paintings, and train the model structure for generating image visual description and feature description;

[0063] S3, using the model of image visual description and feature description to evaluate the painting to be evaluated, obtaining image visual description information and content feature information, and using the language macro model to merge the description information and feature information into an object-centered visual description;

[0064] S4. Use the LLM model to convert the scoring rules into scoring instructions. By providing the LLM model with scoring instructions and image visual description information, the LLM model evaluates the image according to the scoring instructions to obtain the final score of the painting and the reason for the scoring.

[0065] In this implementation, S1 includes:

[0066] S11, collecting training sample painting images;

[0067] S12. Classify and annotate the painting images: image attributes, content description, relationship between content objects, aesthetic evaluation, and emotional evaluation.

[0068] In this implementation, S2 includes:

[0069] S21, grayscale the training sample image, and convert the RGB value of each pixel of the image into a grayscale value of R=G=B:

[0070] Gray(i,j)=0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j);

[0071] Where i, j represent the coordinate position of the pixel, Gray(i, j) is the gray value after processing, R(i, j), G(i, j), B(i, j) represent the RGB values ​​of the pixel respectively;

[0072] S22, performing image enhancement on the training sample image, changing the grayscale distribution of image pixels, enhancing global or local useful information in the image, improving image quality, and enhancing image clarity;

[0073] S23, performing noise reduction processing on the training sample image;

[0074] S24, normalizing the training sample image, changing the size of the image, and changing the image pixel size to a fixed size, the size must be consistent with the requirements of the convolutional neural network;

[0075] S25, extracting features of training sample images;

[0076] S26. Model training. The deep learning framework uses the transform framework and the pre-trained BLIP model with 2.7B parameters to perform multiple rounds of training on the image data after data preprocessing and the image label data to obtain a network model that can quickly return the image visual description information and content feature information. The network model is then saved for subsequent painting scoring testing and operation.

[0077] In this implementation, S23 includes:

[0078] S231, grouping similar blocks, selecting some reference blocks in the noise image, and searching for similar blocks within a certain range near the reference blocks, calculating the similarity between these blocks and the reference blocks, and grouping them according to the similarity:

[0079]

[0080] Where A= ( a1,a2,,a n) and B = ( b1,b2,,b n) are two points in n-dimensional space, d ( A,B ) The Euclidean distance between point A and point B represents the similarity between patches;

[0081] S232, for each grouped three-dimensional matrix, perform two-dimensional transformation using DCT transformation, then perform one-dimensional transformation in the third dimension using Hadamard transformation, perform hard threshold processing on the transformed matrix, set coefficients less than the threshold to 0, and finally obtain the processed image block through inverse transformation;

[0082] S233, combining the processed image blocks into the original image, performing weighted average of the intensity of each pixel by the value of the block at the corresponding position, and obtaining a basic estimated image through the above steps.

[0083] In this implementation manner, the S25 includes:

[0084] S251, acquiring an image and identifying pixel information of an input image, for example, the pixels of the input image are 100*200 pixels;

[0085] S252, image segmentation, image token calculation: the image is divided into multiple 20*20 pixel blocks, the height of the image is divided into 100 / 20=5 row blocks, the width of the image is divided into 200 / 20=10 column blocks, so the total number of blocks is 5*10=50 blocks, each block is flattened into a vector of length 20*20=400, so we have 50 image tokens of length 400;

[0086] S253, linear embedding block, maps each image token of length 400 to an embedding space through a linear transformation, generating an embedding vector of length 768. 768 is the dimension or length of each token after tokenization. After being flattened and linearly transformed, each image block will be mapped to an embedding vector of length 768, which means that the representation of each token is in a 768-dimensional embedding space, so each token is represented as an embedding vector of length 768.

[0087] S254, initialize a token with all 0s as the prediction token;

[0088] S255, concatenate the predicted token with 50 image tokens to form a token sequence. The sequence is in the form of: [ Prediction Token, Image Token1, Image Token2, ..., Image Token50 ] ;

[0089] S256: Add position embedding to all tokens. Position embedding is a vector with the same length as the token sequence, which is used to indicate the position of each token in the sequence.

[0090] S257, image block encoding: normalize the image blocks to ensure that the input data is evenly distributed in different dimensions, use a multi-head self-attention mechanism to capture the relationship between tokens, and obtain the global features of the image. Normalize the data after self-attention processing to ensure that the processed data remains evenly distributed. Finally, use a feedforward neural network: consisting of a fully connected layer, an activation layer, and another fully connected layer to convert the token information. Each token is fully processed and converted to capture the rich feature information of the input image.

[0091] In this implementation, S3 includes:

[0092] S31, preprocessing the painting image, and obtaining a global description, a local description, and a feature description of the image based on the training model generated in S2;

[0093] S32. Use a large language model to merge the global description, local description and feature description of the image into an object-centric visual description.

[0094] In this implementation, S4 includes:

[0095] S41. Through the LLM model, interpret the scoring rules and generate instructions that can be used for scoring. Subsequent scoring uses the scoring instructions.

[0096] S42: Based on the visual description information centered on the image object obtained in S3 and the generated scoring instructions, the powerful reasoning ability of the LLM model is used to perform scoring reasoning one by one according to the instructions to obtain the score and scoring details.

[0097] Embodiment 1:

[0098] The deep learning-based drawing image scoring method and system of the present invention was implemented in School A, a large art training institution located in a certain province in China. School A has a high reputation in the field of art education in the province, especially in the field of youth art training. With the increase in the number of students year by year, the traditional manual scoring method has exposed the problems of low efficiency, strong subjectivity and inconsistent standards when evaluating students' paintings. Especially in large-scale examinations or homework scoring, the manual correction process is cumbersome and prone to errors, which affects the trust of students and parents in the evaluation results.

[0099] In order to meet these challenges, School A decided to introduce the automated scoring system of the present invention, and improve the efficiency, accuracy and interpretability of painting scoring through deep learning technology. During the implementation process, School A first collected student paintings from different grades and majors and their content description label data through the method of the present invention. These data cover multiple attributes of the image, including the color, composition, emotional expression and relationship between objects of the image. The school's technical team preprocessed these data, including grayscale, enhancement, noise reduction and other steps of the image, and trained a deep learning model that can effectively extract image features. In the application, School A applied the scoring system of the present invention to the scoring of a final painting exam. Specifically, during the exam, all the paintings submitted by students were converted into digital images by scanning equipment, and these image data were then standardized by the preprocessing module. The preprocessed image data will be input into the visual description model trained by deep learning, and the model automatically analyzes the content of each work and generates object-centered visual description information. For example, for a work containing characters and background, the system can identify elements such as characters and scenery, and describe their relative positions, light and shadow effects, and overall aesthetic composition.

[0100] During the scoring process, the system first compares the visual description information of the student's painting with the pre-set scoring rules. The scoring rules include requirements for composition, color application, detail processing, creative expression, etc. Through the reasoning ability of the language model, the system can accurately score the works and generate detailed scoring reasons based on each scoring item. For example, for a work with rich colors and reasonable composition, the system will give a high score and explain that "the color matching and composition arrangement conform to the laws of artistic creation and can clearly convey emotions." On the contrary, for works with unreasonable composition, the system will point out that "the proportion between the characters and the background is out of balance, which affects the visual effect of the work." In the application, School A found that the use of the automatic scoring system of the present invention significantly improved the efficiency and consistency of scoring. Taking the final painting exam conducted in December 2019 as an example, a total of 500 students submitted paintings. When using the traditional manual scoring method, teachers need to spend 30 minutes to review each work, and the entire scoring process takes about 250 hours in total. After using the automatic scoring system of the present invention, the whole process only takes 4 hours, saving more than 90% of the time. More importantly, the scoring reasons generated by the system are clear and transparent, which can eliminate students' and parents' doubts about the scoring and improve the fairness and acceptability of the scoring.

[0101] Through comparative analysis, School A also found that the consistency of scoring results has been significantly improved after adopting the automatic scoring system. Traditional manual scoring often leads to large fluctuations in scoring due to different subjective factors of the scorers, while the automatic scoring system based on deep learning has good consistency in scoring results for all works, and the scoring error is significantly reduced. Specifically, under the same scoring criteria, the error range of manual scoring is ±15 points, while the error range of system scoring is reduced to ±3 points, and the consistency of scoring has increased by about 5 times.

[0102] Table 1 Comparison of data before and after painting evaluation

[0103]

[0104] As shown in Table 1 above, the automatic scoring system of the present invention not only improves scoring efficiency, but also effectively improves the consistency and transparency of scoring results. In this way, School A can efficiently complete the evaluation of a large number of student works in a short period of time and ensure that each scoring item has a reasonable basis and explanation.

[0105] The system can also automatically identify and count the elements in the image that do not meet the scoring criteria, and list them in detail when generating the scoring report. For example, if the color usage of the work is unreasonable, the system will specifically point out which color blocks do not match the aesthetic rules and give suggestions for improvement. These functions greatly enhance the interpretability of the system, allowing students and parents to more clearly understand the basis behind the scoring results.

[0106] After this implementation, School A successfully solved the subjective bias and inefficiency problems that occurred in previous manual scoring, providing a more intelligent and efficient solution for the school's art education management. With the application of the system in more courses and scenarios, the teaching quality of School A has been further improved, and it has also provided valuable experience for the digital transformation of other educational institutions.

[0107] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A drawing image scoring method and system based on deep learning, characterized in that: include: The preprocessing module is used to perform preliminary processing on the input image, including grayscale, image enhancement, denoising, normalization, and feature extraction to ensure the accuracy of subsequent analysis; Model training is used to perform deep learning training on various painting images and image description label data to obtain an evaluation model for image content description, which is mainly responsible for converting images into multi-level, object-centric visual descriptions; The scoring rule conversion module is used to convert the scoring rules input by the user into a series of evaluation instructions; The scoring module uses large language models to measure the alignment between the generated image and text based on the obtained image visual description, and uses their reasoning ability to understand the combined properties of objects and the complex interactions between multiple objects. Based on the above generated visual description, the entire evaluation process consists of three steps: instruction follow-up scoring, rule-enhanced scoring, and scoring reason generation; The output module performs weighted calculations on the scores of each instruction based on the scoring instructions, and finally outputs the scores of the students' paintings and the reasons for the scores.

2. A drawing image scoring method based on deep learning according to claim 1, characterized in that: The following steps are involved: S1, collect painting images and image content description label training samples; S2. Perform grayscale conversion, image enhancement, noise reduction, normalization, and image feature extraction on the training paintings, and train the model structure for generating image visual description and feature description; S3, using the model of image visual description and feature description to evaluate the painting to be evaluated, obtaining image visual description information and content feature information, and using the language macro model to merge the description information and feature information into an object-centered visual description; S4. Use the LLM model to convert the scoring rules into scoring instructions. By providing the LLM model with scoring instructions and image visual description information, the LLM model evaluates the image according to the scoring instructions to obtain the final score of the painting and the reason for the scoring.

3. The automatic painting scoring method based on deep learning according to claim 2 is characterized in that: The S1 includes: S11, collecting training sample painting images; S12. Classify and annotate the painting images: image attributes, content description, relationship between content objects, aesthetic evaluation, and emotional evaluation.

4. The drawing image scoring method based on deep learning according to claim 2, characterized in that: The S2 includes: S21, grayscale the training sample image, and convert the RGB value of each pixel of the image into a grayscale value of R=G=B: Gray(i,j)=0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j); Where i, j represent the coordinate position of the pixel, Gray(i, j) is the gray value after processing, R(i, j), G(i, j), B(i, j) represent the RGB values ​​of the pixel respectively; S22, performing image enhancement on the training sample image, changing the grayscale distribution of image pixels, enhancing global or local useful information in the image, improving image quality, and enhancing image clarity; S23, performing noise reduction processing on the training sample image; S24, normalizing the training sample image, changing the size of the image, and changing the image pixel size to a fixed size, the size must be consistent with the requirements of the convolutional neural network; S25, extracting features of training sample images; S26. Model training. The deep learning framework uses the transform framework and the pre-trained BLIP model with 2.7B parameters to perform multiple rounds of training on the image data after data preprocessing and the image label data to obtain a network model that can quickly return the image visual description information and content feature information. The network model is then saved for subsequent painting scoring testing and operation.

5. A drawing image scoring method based on deep learning according to claim 4, characterized in that: The S23 includes: S231, grouping similar blocks, selecting some reference blocks in the noise image, and searching for similar blocks within a certain range near the reference blocks, calculating the similarity between these blocks and the reference blocks, and grouping them according to the similarity: Where A= ( a1,a2,,a n) and B = ( b1,b2,,b n) are two points in n-dimensional space, d ( A,B ) The Euclidean distance between point A and point B represents the similarity between patches; S232, for each grouped three-dimensional matrix, perform two-dimensional transformation using DCT transformation, then perform one-dimensional transformation in the third dimension using Hadamard transformation, perform hard threshold processing on the transformed matrix, set coefficients less than the threshold to 0, and finally obtain the processed image block through inverse transformation; S233, combining the processed image blocks into the original image, performing weighted average of the intensity of each pixel by the value of the block at the corresponding position, and obtaining a basic estimated image through the above steps.

6. A drawing image scoring method based on deep learning according to claim 4, characterized in that: The S25 includes: S251, acquiring an image and identifying pixel information of an input image, for example, the pixels of the input image are 100*200 pixels; S252, image segmentation, image token calculation: the image is divided into multiple 20*20 pixel blocks, the height of the image is divided into 100 / 20=5 row blocks, the width of the image is divided into 200 / 20=10 column blocks, so the total number of blocks is 5*10=50 blocks, each block is flattened into a vector of length 20*20=400, so we have 50 image tokens of length 400; S253, linear embedding block, maps each image token of length 400 to an embedding space through a linear transformation, generating an embedding vector of length 768. 768 is the dimension or length of each token after tokenization. After being flattened and linearly transformed, each image block will be mapped to an embedding vector of length 768, which means that the representation of each token is in a 768-dimensional embedding space, so each token is represented as an embedding vector of length 768. S254, initialize a token with all 0s as the prediction token; S255, concatenate the predicted token with 50 image tokens to form a token sequence, the sequence format is: [predicted token, image token1, image token2, ..., image token50]; S256: Add position embedding to all tokens. Position embedding is a vector with the same length as the token sequence, which is used to indicate the position of each token in the sequence. S257, image block encoding: normalize the image blocks to ensure that the input data is evenly distributed in different dimensions, use a multi-head self-attention mechanism to capture the relationship between tokens, and obtain the global features of the image. Normalize the data after self-attention processing to ensure that the processed data remains evenly distributed. Finally, use a feedforward neural network: consisting of a fully connected layer, an activation layer, and another fully connected layer to convert the token information. Each token is fully processed and converted to capture the rich feature information of the input image.

7. The method for scoring a drawing image based on deep learning according to claim 2, characterized in that: The S3 includes: S31, preprocessing the painting image, and obtaining a global description, a local description, and a feature description of the image based on the training model generated in S2; S32. Use a large language model to merge the global description, local description and feature description of the image into an object-centric visual description.

8. The method for scoring a drawing image based on deep learning according to claim 2, characterized in that: The S4 includes: S41. Through the LLM model, interpret the scoring rules and generate instructions that can be used for scoring. Subsequent scoring uses the scoring instructions. S42: Based on the visual description information centered on the image object obtained in S3 and the generated scoring instructions, the powerful reasoning ability of the LLM model is used to perform scoring reasoning one by one according to the instructions to obtain the score and scoring details.

Images