Digital art work intelligent quality detection and optimization recommendation system

Abstract

The application relates to a digital art work intelligent quality detection and optimization recommendation system, and relates to the technical field of digital art. The application converts expert perceptual aesthetics into a quantitative operator through the cooperation of the artificial auxiliary judgment and the refined classification unit, overcomes the limitation that a traditional algorithm cannot understand a small number of factions and complex techniques, significantly improves the delicacy of classification and the objectivity of evaluation, simultaneously introduces active learning, only needs a small amount of manual marking of edge cases, greatly reduces the labor cost and improves the classification confidence, can accurately identify AI disguises from the dual dimensions of pixel-level traces and semantic logic, provides a quantitative basis for originality judgment, generates a customized report through a dynamic weight matrix, accurately matches an optimization rule library, provides a substantial auxiliary optimization means for creators, and has very high industrial practical value.

Description

Technical Field

[0001] This invention relates to the field of digital art technology, and in particular to an intelligent quality detection and optimization recommendation system for digital artworks. Background Technology

[0002] In the field of digital art, although intelligent quality inspection has made significant progress, existing inspection systems still face many deep-seated problems in the face of the uniqueness and complexity of artistic creation.

[0003] I. The Contradiction Between Subjectivity and Objectivity in Aesthetic Evaluation: Intelligent systems are typically trained on massive amounts of data using deep learning models to learn popular aesthetics. Algorithms tend to provide average aesthetic scores, making it difficult to understand the artistic value of specific styles such as the beauty of the ugly or minimalism, or to grasp the cultural symbolism, historical metaphors, or social judgments behind art. Existing AI struggles to perceive the human traces or emotional tension inherent in artworks, often simplifying them to physical parameters such as composition and color distribution.

[0004] II. Challenges in Detecting AIGC-Generated Content: With the popularization of AI paintings, quality detection systems are encountering new problems when dealing with such works: many AI works are visually striking, but upon closer inspection, logical flaws are apparent. Current detection models remain relatively weak in spatial logical reasoning. For realistic works, subtle unnatural textures can lead to visual discomfort, and existing quality detection metrics struggle to capture this psychological discomfort.

[0005] Third, the lack of judgment on stylistics and originality: Intelligent detection can identify visual similarity, but it is difficult to define the legal and artistic boundaries between stylistic borrowing and low-level imitation. Evaluation algorithms often give high scores to works that conform to conventional compositions, thereby suppressing artworks that break the rules and have subversive creations;

[0006] To address the aforementioned technical deficiencies, a solution is proposed. Summary of the Invention

[0007] The purpose of this invention is to transform expert aesthetic judgment into quantitative operators through the synergy of human-assisted judgment and refined classification units. This overcomes the limitations of traditional algorithms in understanding niche schools of thought and complex techniques, significantly improving the finesse of classification and the objectivity of evaluation. At the same time, by introducing active learning, only a very small amount of human intervention is needed to label marginal cases, greatly reducing labor costs and improving classification confidence.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent quality detection and optimization recommendation system for digital artworks, comprising a human-assisted judgment unit, a refined classification unit, a feature clustering correction unit, an AIGC recognition unit, and a comprehensive evaluation unit, wherein:

[0009] The human-assisted judgment unit is used to provide a preset human interface to obtain digital artworks and receive art feature parameters set in advance by the art appreciation team. The digital artworks are initially classified by humans, and the classification results are sent to the refined classification unit after being labeled with auxiliary annotations.

[0010] The refined classification unit is used to obtain the art feature parameters set by the art appreciation team. The art feature parameters are the core feature operators of a specific art school defined by experts. A digital art evaluation model is constructed using a convolutional neural network. The model is trained based on the core feature operators. The digital art works are further classified into schools and techniques. The refined classification results are then labeled on the digital art works and sent to the feature clustering correction unit.

[0011] The feature clustering correction unit is used to obtain the refined classification results of digital artworks. After the features are initially extracted based on the manifold learning algorithm, the edge cases are quickly labeled by manual means, and the digital art evaluation model is updated based on the active learning mechanism to obtain the extremely fine-grained classification results of digital artworks, which are then sent to the comprehensive evaluation unit.

[0012] The AIGC recognition unit is used to acquire digital artworks and extract feature fingerprints from them. By comparing them with a preset AI-generated feature library, it outputs the AI ​​creation probability and recognition score and sends them to the comprehensive evaluation unit.

[0013] The comprehensive evaluation unit combines the fine-grained classification results with the recognition score, calls the quality evaluation weight matrix under the corresponding category, generates a comprehensive quality report, and, based on the deduction items in the comprehensive quality report, matches a preset art optimization rule library and outputs a recommended modification strategy.

[0014] Furthermore, the specific process of manually performing preliminary classification of digital artworks is as follows:

[0015] S101. Access the knowledge base of the art appreciation team through a pre-set interface. The art appreciation team sets a set of art feature parameters based on color psychology, composition principles and historical school characteristics. The art feature parameters include color tendency, composition form and brushstroke texture characteristics.

[0016] S102. After obtaining the digital artwork to be tested, the digital artwork is displayed to the art appreciation team through an interactive panel. The appraisers select from the preset macro categories based on their visual perception to obtain the primary classification labels.

[0017] S103. After completing the initial classification, the appreciation team uses the interactive panel to assist in annotating the key aesthetic anchors in the digital artworks. They manually select areas that are representative of different styles on the digital artworks, convert the selected areas into pixel coordinates, and attach corresponding label text to obtain key area coordinate labels.

[0018] S104. Encapsulate the digital artwork, primary classification labels, and key area coordinate labels into an initial feature package, and send the encapsulated initial feature package to the refined classification unit.

[0019] Furthermore, the specific process of constructing a digital art evaluation model and further categorizing digital artworks by genre and technique is as follows:

[0020] S201. Obtain the core feature operators of a specific art style as defined by experts, and transform the core feature operators defined by experts into tensor forms that can be processed by computers.

[0021] S202. A digital art evaluation model is constructed using a convolutional neural network architecture with an attention mechanism. The digital art evaluation model is used to capture the macroscopic composition and layout features of the artwork, and at the same time, it extracts the microscopic details of texture and brushstrokes through wavelet transform to distinguish between hand-painted digital art and textures generated by pure algorithms.

[0022] S203. Using art feature parameters as training samples, transfer learning is performed on the pre-trained model, and an operator consistency regularization term is introduced into the loss function. That is, the classification probability output by the digital art evaluation model conforms to the label, and the intermediate layer feature mapping extracted by it is consistent with the feature operator defined by the expert.

[0023] S204. After inputting the digital artwork to be detected into the digital art evaluation model, perform multi-layer convolution processing on the digital artwork and generate strong activation in specific neuron layers according to the preset operator.

[0024] S205. A hierarchical classification path is adopted to output a refined classification result, wherein the refined classification result includes the first level of genre identification and the second level of technique decomposition.

[0025] S206. The final refined classification results are converted into structured data in XML format. The structured data is used as re-annotation information, superimposed on the feature stream of the original digital artwork, and sent to the feature clustering correction unit.

[0026] Furthermore, the specific process for obtaining extremely fine-grained classification results of digital artworks is as follows:

[0027] S301. Obtain detailed classification results of digital artworks, and use manifold learning to project the high-dimensional artistic features of digital artworks into a low-dimensional space for clustering.

[0028] S302. Automatically identify digital artworks located at the edge of the cluster center and multiple cluster boundaries, determine them as edge cases, call the uncertainty sampling algorithm in the active learning strategy, filter out the top N samples with the lowest model prediction confidence, and then display the top N samples to the artwork appreciation team through the interactive panel.

[0029] S303. The classification labels in the samples are corrected by the appraisers, and the specific local features that cause confusion in the digital art evaluation model (such as special texture brushstrokes and light and shadow transitions) are semantically enhanced and labeled.

[0030] S304. Add manually labeled samples to the core training set to trigger incremental training of the digital art evaluation model. By comparing loss functions, the distance between works belonging to the same fine-grained category is narrowed, and the distance between easily confused categories is widened. The feature space is re-segmented to form multi-dimensional combination of fine-grained classification labels.

[0031] S305. The revised fine-grained classification results include: school affiliation, technique combination, aesthetic preference, and classification confidence. The fine-grained classification results are encapsulated and sent to the comprehensive evaluation unit as the unique index for subsequent calls to the weight matrix.

[0032] Furthermore, the specific process for outputting the AI ​​creation probability and recognition score is as follows:

[0033] S401. After acquiring the digital artwork to be detected, use a high-pass filter to extract the high-frequency residual components in the digital artwork, and at the same time extract the semantic feature fingerprint of the digital artwork.

[0034] S402. Input the extracted feature fingerprint into a preset AI-generated feature library, wherein the AI-generated feature library includes:

[0035] A model signature library used to record specific convolutional artifacts generated by different versions of the StableDiffusion model;

[0036] A physical contradiction feature library, used for AI's common non-natural light and shadow contradictions and non-Euclidean spatial structure feature templates;

[0037] S403. Using cosine similarity or Euclidean distance, calculate the degree of matching between the feature fingerprint of the digital artwork to be detected and typical AI features in the AI-generated feature library.

[0038] S404. A weighted fusion algorithm is used to calculate the initial probability by combining high-frequency noise similarity and semantic logic contradiction, and the AI ​​creation probability value is output by combining the extremely fine-grained classification results.

[0039] S405. Map the AI ​​creation probability value to a preset rating scale to obtain a recognition score.

[0040] Furthermore, the specific process for generating a comprehensive quality report and outputting recommended modification strategies is as follows:

[0041] S501. Obtain the extremely fine-grained classification results from the feature clustering correction unit, and retrieve and call the corresponding exclusive weight matrix M from the preset database according to the classification label;

[0042] S502. Obtain the recognition score output by the AIGC recognition unit and the basic quality parameters Fi of the digital artwork (such as contrast, information entropy, and color harmony), and calculate the final evaluation score Q according to the following formula:

[0043]

[0044] Where Fi represents the visual feature values ​​and Wi represents the corresponding weights in matrix M;

[0045] The function is a modulatory function for the AI ​​creation probability Pi based on the classification context C, used to determine the originality contribution of the work.

[0046] S503. The evaluation score includes technical score, aesthetic score, originality score and comprehensive total score. After integration, a comprehensive quality report is generated. The low-scoring items in the comprehensive quality report are scanned, and the feature vectors corresponding to the deducted items are extracted. If the AIGC recognition unit reports that there are high-frequency logical contradictions in the area corresponding to the deducted items, then it is determined to be a key defect point.

[0047] S504. Input the key defect points into a preset art optimization rule base, which contains targeted suggestions preset by experts:

[0048] If points are deducted due to AI logic errors, a suggestion will be provided to manually restructure the structure and correct the unnatural deformation of key defects.

[0049] If points are deducted due to color disharmony, the output should be shifted based on the color wheel rules of the corresponding school of thought to enhance color expressiveness.

[0050] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0051] This intelligent quality detection and optimization recommendation system for digital artworks transforms expert aesthetic judgment into quantitative operators through the collaboration of human-assisted judgment and refined classification units. It overcomes the limitations of traditional algorithms in understanding niche genres and complex techniques, significantly improving the granularity of classification and the objectivity of evaluation. At the same time, it introduces active learning, requiring only a minimal amount of human intervention to label marginal cases, greatly reducing labor costs and increasing classification confidence. It can accurately detect AI disguises from both pixel-level traces and semantic logic dimensions, providing a quantitative basis for originality judgment. It generates customized reports through dynamic weight matrices and accurately matches optimization rule bases, providing creators with substantial auxiliary optimization methods and possessing extremely high industrial practical value. Attached Figure Description

[0052] Figure 1 A schematic diagram of the overall structure of the present invention is shown. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example:

[0055] like Figure 1 As shown, a digital art intelligent quality detection and optimization recommendation system includes a human-assisted judgment unit, a refined classification unit, a feature clustering correction unit, an AIGC recognition unit, and a comprehensive evaluation unit, wherein:

[0056] The human-assisted judgment unit is used to provide a preset human interface to obtain digital artworks and receive art feature parameters set in advance by the art appreciation team. The digital artworks are initially classified by humans, and the classification results are sent to the refined classification unit after being labeled with auxiliary annotations.

[0057] The specific process of manually classifying digital artworks into preliminary categories is as follows:

[0058] S101. Access the knowledge base of the art appreciation team through a pre-set interface. The art appreciation team sets a set of art feature parameters based on color psychology, composition principles and historical school characteristics. The art feature parameters include color tendency (warm and cool, saturation distribution), composition form (rule of thirds, golden ratio, centripetal), and brushstroke texture characteristics (flat wash, smudged, thick wash).

[0059] S102. After obtaining the digital artwork to be tested, the digital artwork is displayed to the art appreciation team through an interactive panel. The appraisers select from the preset macro categories (such as: realistic style, abstract style, vector graphics, pixel art, etc.) based on their visual perception to obtain the primary classification labels.

[0060] S103. After completing the initial classification, the appreciation team uses the interactive panel to assist in annotating the key aesthetic anchors in the digital artworks. They manually select areas that are representative of different styles on the digital artworks, convert the selected areas into pixel coordinates, and attach corresponding label text to obtain key area coordinate labels.

[0061] S104. Encapsulate the digital artwork, primary classification labels, and key area coordinate labels into an initial feature package, and send the encapsulated initial feature package to the refined classification unit.

[0062] The refined classification unit is used to obtain the art feature parameters set by the art appreciation team. The art feature parameters are specifically the core feature operators of a specific art school defined by experts. A digital art evaluation model is constructed using a convolutional neural network. The model is trained based on the core feature operators. The digital artworks are further classified into schools and techniques. The refined classification results are then labeled on the digital artworks and sent to the feature clustering correction unit.

[0063] The specific process of constructing a digital art evaluation model and further classifying digital artworks by genre and technique is as follows:

[0064] S201. Obtain the core feature operators of a specific art movement defined by experts, and transform the core feature operators defined by experts into tensor forms that can be processed by computers. For example, for the "Impressionism" operator, its corresponding technical features include: the specific distribution of the color contrast histogram, the Fourier transform features of the brushstroke frequency, etc. These operators serve as supervision signals to guide the feature selection of the convolutional neural network (CNN), ensuring that the model not only learns the low-level pixel information, but also learns the high-level artistic logic.

[0065] S202. A digital art evaluation model is constructed using a convolutional neural network architecture with an attention mechanism. The digital art evaluation model is used to capture the macroscopic composition and layout features of the artwork, and at the same time, it extracts the microscopic details of texture and brushstrokes through wavelet transform to distinguish between hand-painted digital art and textures generated by pure algorithms.

[0066] S203. Using art feature parameters as training samples, transfer learning is performed on the pre-trained model, and an operator consistency regularization term is introduced into the loss function. That is, the classification probability output by the digital art evaluation model conforms to the label, and the intermediate layer feature mapping extracted by it is consistent with the feature operator defined by the expert.

[0067] S204. After inputting the digital artwork to be tested into the digital art evaluation model, the digital artwork is subjected to multi-layer convolution processing. According to the preset operator, strong activation is generated in a specific neuron layer. For example, if the artwork has strong geometric segmentation features, the corresponding "Cubist operator" neuron group will output a high value.

[0068] S205. A hierarchical classification path is adopted to output a refined classification result, wherein the refined classification result includes the first level of genre identification and the second level of technique decomposition.

[0069] The first level (style identification): Determining it as a major category, such as "modernism".

[0070] The second layer (technique breakdown): Further analysis under the major category, such as determining it as "geometric abstraction" or "lyrical abstraction".

[0071] S206. The final refined classification results (e.g., [genre: Pop Art; technique: screen printing simulation; color mode: high contrast color]) are converted into structured data in XML format. The structured data is superimposed on the feature stream of the original digital artwork as re-annotation information and sent to the feature clustering correction unit.

[0072] The feature clustering correction unit is used to obtain the refined classification results of digital artworks. After the features are initially extracted based on the manifold learning algorithm, the edge cases are quickly labeled by manual means, and the digital art evaluation model is updated based on the active learning mechanism to obtain the extremely fine-grained classification results of digital artworks, which are then sent to the comprehensive evaluation unit.

[0073] The specific process for obtaining extremely fine-grained classification results of digital artworks is as follows:

[0074] S301. Obtain detailed classification results of digital artworks, and use manifold learning to project the high-dimensional artistic features of digital artworks into a low-dimensional space for clustering.

[0075] S302. Automatically identify digital artworks located at the edge of the cluster center and multiple cluster boundaries, determine them as edge cases, call the uncertainty sampling algorithm in the active learning strategy, filter out the top N samples with the lowest model prediction confidence, and then display the top N samples to the artwork appreciation team through the interactive panel.

[0076] S303. The classification labels in the samples are corrected by the appraisers, and the specific local features that cause confusion in the digital art evaluation model (such as special texture brushstrokes and light and shadow transitions) are semantically enhanced and labeled.

[0077] S304. Add manually labeled samples to the core training set to trigger incremental training of the digital art evaluation model. By comparing loss functions, the distance between works belonging to the same fine-grained category is narrowed, and the distance between easily confused categories is widened. The feature space is re-segmented to form a multi-dimensional combination of fine-grained classification labels such as "cyberpunk style - neon cool color - high contrast - realistic brushstrokes".

[0078] S305. The revised fine-grained classification results include: school affiliation, technique combination, aesthetic preference, and classification confidence. The fine-grained classification results are encapsulated and sent to the comprehensive evaluation unit as the unique index for subsequent calls to the weight matrix.

[0079] The AIGC recognition unit is used to acquire digital artworks and extract feature fingerprints from them. By comparing them with a preset AI-generated feature library, it outputs the AI ​​creation probability and recognition score and sends them to the comprehensive evaluation unit.

[0080] The specific process for outputting the AI ​​creation probability and recognition score is as follows:

[0081] S401. After acquiring the digital artwork to be detected, use a high-pass filter to extract the high-frequency residual components in the digital artwork, and at the same time extract the semantic feature fingerprint of the digital artwork, focusing on the coherence of local details, such as the disappearance of the end of the brushstroke, the overflow of color at the edge of the object, and the topological logic of complex structures (such as hands, fences, and text).

[0082] S402. Input the extracted feature fingerprint into a preset AI-generated feature library, wherein the AI-generated feature library includes:

[0083] A model signature library used to record specific convolutional artifacts generated by different versions of the StableDiffusion model;

[0084] A physical contradiction feature library, used for AI's common non-natural light and shadow contradictions and non-Euclidean spatial structure feature templates;

[0085] S403. Using cosine similarity or Euclidean distance, calculate the degree of matching between the feature fingerprint of the digital artwork to be detected and typical AI features in the AI-generated feature library.

[0086] S404. A weighted fusion algorithm is used to calculate the initial probability by combining high-frequency noise similarity and semantic logic contradiction, and the AI ​​creation probability value is output by combining the extremely fine-grained classification results.

[0087] For example, if the classification result is "Generative Art (Generative Art)", the penalty for algorithmic traces is relaxed; if the classification result is "Traditional Digital Painting", the weight of logically contradictory terms is increased.

[0088] S405. Map the AI ​​creation probability value to a preset rating scale to obtain a recognition score.

[0089] The recognition score is not only a total score, but also includes a generation trace score (reflecting the degree of technical generation) and a logical flaw score (reflecting the degree of low quality generated by AI).

[0090] The comprehensive evaluation unit combines the fine-grained classification results with the recognition score, calls the quality evaluation weight matrix under the corresponding category, generates a comprehensive quality report, and, based on the deduction items in the comprehensive quality report, matches a preset art optimization rule library and outputs a recommended modification strategy.

[0091] The specific process for generating a comprehensive quality report and outputting recommended modification strategies is as follows:

[0092] S501. Obtain the extremely fine-grained classification results from the feature clustering correction unit, and retrieve and call the corresponding exclusive weight matrix M from the preset database according to the classification label;

[0093] S502. Obtain the recognition score output by the AIGC recognition unit and the basic quality parameters Fi of the digital artwork (such as contrast, information entropy, and color harmony), and calculate the final evaluation score Q according to the following formula:

[0094]

[0095] Where Fi represents the visual feature values ​​and Wi represents the corresponding weights in matrix M;

[0096] The function is a modulatory function for the AI ​​creation probability Pi based on the classification context C, used to determine the originality contribution of the work.

[0097] S503. The evaluation score includes technical score, aesthetic score, originality score and comprehensive total score. After integration, a comprehensive quality report is generated. The low-scoring items in the comprehensive quality report are scanned, and the feature vectors corresponding to the deducted items are extracted. If the AIGC recognition unit reports that there are high-frequency logical contradictions in the area corresponding to the deducted items, then it is determined to be a key defect point.

[0098] S504. Input the key defect points into a preset art optimization rule base, which contains targeted suggestions preset by experts:

[0099] If points are deducted due to AI logic errors, a suggestion will be provided to manually restructure the structure and correct the unnatural deformation of key defects.

[0100] If points are deducted due to color disharmony, the output should be shifted based on the color wheel rules of the corresponding school of thought to enhance color expressiveness.

[0101] This invention transforms expert aesthetic judgment into quantitative operators through the synergy of human-assisted judgment and refined classification units. This overcomes the limitations of traditional algorithms in understanding niche genres and complex techniques, significantly improving the granularity of classification and the objectivity of evaluation. Simultaneously, it introduces active learning, requiring only minimal human annotation of marginal cases, greatly reducing labor costs and increasing classification confidence. It can accurately detect AI disguises from both pixel-level traces and semantic logic dimensions, providing a quantitative basis for originality assessment. Furthermore, it generates customized reports through a dynamic weight matrix and accurately matches and optimizes the rule base, providing creators with substantial auxiliary optimization methods and possessing extremely high industrial practical value.

[0102] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0103] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0104] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital artwork intelligent quality detection and optimization recommendation system, characterized in that, It includes a human-assisted judgment unit, a refined classification unit, a feature clustering correction unit, an AIGC recognition unit, and a comprehensive evaluation unit, among which: The human-assisted judgment unit is used to provide a preset human interface to obtain digital artworks and receive art feature parameters set in advance by the art appreciation team. The digital artworks are initially classified by humans, and the classification results are sent to the refined classification unit after being labeled with auxiliary annotations. The specific process of manually classifying digital artworks into preliminary categories is as follows: S101. Access the knowledge base of the art appreciation team through a pre-set interface. The art appreciation team sets a set of art feature parameters based on color psychology, composition principles and historical school characteristics. The art feature parameters include color tendency, composition form and brushstroke texture characteristics. S102. After obtaining the digital artwork to be tested, the digital artwork is displayed to the art appreciation team through an interactive panel. The appraisers select from the preset macro categories based on their visual perception to obtain the primary classification labels. S103. After completing the initial classification, the appreciation team uses the interactive panel to assist in annotating the key aesthetic anchors in the digital artworks. They manually select areas that are representative of different styles on the digital artworks, convert the selected areas into pixel coordinates, and attach corresponding label text to obtain key area coordinate labels. S104. Encapsulate the digital artwork, primary classification labels, and key area coordinate labels into an initial feature package, and send the encapsulated initial feature package to the refined classification unit. The refined classification unit is used to obtain the art feature parameters set by the art appreciation team. The art feature parameters are specifically the core feature operators of a specific art school defined by experts. A digital art evaluation model is constructed using a convolutional neural network. The model is trained based on the core feature operators. The digital artworks are further classified into schools and techniques. The refined classification results are then labeled on the digital artworks and sent to the feature clustering correction unit. The feature clustering correction unit is used to obtain the refined classification results of digital artworks. After the features are initially extracted based on the manifold learning algorithm, the edge cases are quickly labeled by manual means, and the digital art evaluation model is updated based on the active learning mechanism to obtain the extremely fine-grained classification results of digital artworks, which are then sent to the comprehensive evaluation unit. The specific process for obtaining extremely fine-grained classification results of digital artworks is as follows: S301. Obtain detailed classification results of digital artworks, and use manifold learning to project the high-dimensional artistic features of digital artworks into a low-dimensional space for clustering. S302. Automatically identify digital artworks located at the edge of the cluster center and multiple cluster boundaries, determine them as edge cases, call the uncertainty sampling algorithm in the active learning strategy, filter out the top N samples with the lowest model prediction confidence, and then display the top N samples to the artwork appreciation team through the interactive panel. S303. The classification labels in the samples are corrected by the appraisers, and the specific local features that cause confusion in the digital art evaluation model are semantically enhanced and labeled. S304. Add manually labeled samples to the core training set to trigger incremental training of the digital art evaluation model. By comparing loss functions, the distance between works belonging to the same fine-grained category is narrowed, and the distance between easily confused categories is widened. The feature space is re-segmented to form multi-dimensional combination of fine-grained classification labels. S305. The revised fine-grained classification results include: school affiliation, technique combination, aesthetic preference, and classification confidence. The fine-grained classification results are encapsulated and sent to the comprehensive evaluation unit as the unique index for subsequent calls to the weight matrix. The AIGC recognition unit is used to acquire digital artworks and extract feature fingerprints from them. By comparing them with a preset AI-generated feature library, it outputs the AI ​​creation probability and recognition score and sends them to the comprehensive evaluation unit. The comprehensive evaluation unit combines the fine-grained classification results with the recognition score, calls the quality evaluation weight matrix under the corresponding category, generates a comprehensive quality report, and, based on the deduction items in the comprehensive quality report, matches a preset art optimization rule library and outputs a recommended modification strategy.

2. The intelligent quality detection and optimization recommendation system for digital artworks according to claim 1, characterized in that, The specific process of constructing a digital art evaluation model and further classifying digital artworks by genre and technique is as follows: S201. Obtain the core feature operators of a specific art style as defined by experts, and transform the core feature operators defined by experts into tensor forms that can be processed by computers. S202. A digital art evaluation model is constructed using a convolutional neural network architecture with an attention mechanism. The digital art evaluation model is used to capture the macroscopic composition and layout features of the artwork, and at the same time, it extracts the microscopic details of texture and brushstrokes through wavelet transform to distinguish between hand-painted digital art and textures generated by pure algorithms. S203. Using art feature parameters as training samples, transfer learning is performed on the pre-trained model, and an operator consistency regularization term is introduced into the loss function. That is, the classification probability output by the digital art evaluation model conforms to the label, and the intermediate layer feature mapping extracted by it is consistent with the feature operator defined by the expert. S204. After inputting the digital artwork to be detected into the digital art evaluation model, perform multi-layer convolution processing on the digital artwork and generate strong activation in specific neuron layers according to the preset operator. S205. A hierarchical classification path is adopted to output a refined classification result, wherein the refined classification result includes the first level of genre identification and the second level of technique decomposition. S206. The final refined classification results are converted into structured data in XML format. The structured data is used as re-annotation information, superimposed on the feature stream of the original digital artwork, and sent to the feature clustering correction unit.

3. The intelligent quality detection and optimization recommendation system for digital artworks according to claim 1, characterized in that, The specific process for outputting the AI ​​creation probability and recognition score is as follows: S401. After acquiring the digital artwork to be detected, use a high-pass filter to extract the high-frequency residual components in the digital artwork, and at the same time extract the semantic feature fingerprint of the digital artwork. S402. Input the extracted feature fingerprint into a preset AI-generated feature library, wherein the AI-generated feature library includes: A model signature library used to record specific convolutional artifacts generated by different versions of the StableDiffusion model; A physical contradiction feature library, used for AI's common non-natural light and shadow contradictions and non-Euclidean spatial structure feature templates; S403. Using cosine similarity or Euclidean distance, calculate the degree of matching between the feature fingerprint of the digital artwork to be detected and typical AI features in the AI-generated feature library. S404. A weighted fusion algorithm is used to calculate the initial probability by combining high-frequency noise similarity and semantic logic contradiction, and the AI ​​creation probability value is output by combining the extremely fine-grained classification results. S405. Map the AI ​​creation probability value to a preset rating scale to obtain a recognition score.

4. The intelligent quality detection and optimization recommendation system for digital artworks according to claim 1, characterized in that, The specific process for generating a comprehensive quality report and outputting recommended modification strategies is as follows: S501. Obtain the extremely fine-grained classification results from the feature clustering correction unit, and retrieve and call the corresponding exclusive weight matrix M from the preset database according to the classification label; S502. Obtain the recognition score output by the AIGC recognition unit and the basic quality parameters Fi of the digital artwork, and calculate the final evaluation score Q according to the following formula: Where Fi represents the visual feature values ​​and Wi represents the corresponding weights in matrix M; The function is a modulatory function for the AI ​​creation probability Pi based on the classification context C, used to determine the originality contribution of the work. S503. The evaluation score includes technical score, aesthetic score, originality score and comprehensive total score. After integration, a comprehensive quality report is generated. The low-scoring items in the comprehensive quality report are scanned, and the feature vectors corresponding to the deducted items are extracted. If the AIGC recognition unit reports that there are high-frequency logical contradictions in the area corresponding to the deducted items, then it is determined to be a key defect point. S504. Input the key defect points into a preset art optimization rule base, which contains targeted suggestions preset by experts: If points are deducted due to AI logic errors, a suggestion will be provided to manually restructure the structure and correct the unnatural deformation of key defects. If points are deducted due to color disharmony, the output should be shifted based on the color wheel rules of the corresponding school of thought to enhance color expressiveness.

Images