Paper-cut screening method and device based on target detection and collaborative evaluation, and medium

Abstract

The present application belongs to the field of artificial intelligence and computer vision technology, and specifically relates to a paper-cut screening method and device based on target detection and collaborative evaluation and a medium. The present application controls the cross-modal condition and fine-tunes the LoRA parameters of the standardized paper-cut materials, generates paper-cut images conforming to specific cultural themes by using a Flux.1 diffusion model; subsequently, an improved YOLOv11n-PPA lightweight target detection network is innovatively constructed, partial convolution architecture is combined with a global attention mechanism to accurately detect and eliminate the tiny defects such as line breakage and adhesion in the generated images, and an objective candidate image set with high process standards is output; finally, a fuzzy comprehensive evaluation model based on the core weight of process feasibility is used to perform secondary subjective calculation and sorting on the candidate image set, and the optimal paper-cut image is output. The present application breaks through the whole-process evaluation closed loop from image generation, defect filtering to aesthetic optimization, and significantly improves the efficiency and quality of non-heritage paper-cut cultural and creative transformation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer vision technology, and in particular to a paper-cutting screening method, device and medium based on target detection and collaborative evaluation. Background Technology

[0002] Paper cutting, as an outstanding representative of Chinese intangible cultural heritage, carries the profound cultural spirit and artistic wisdom of various ethnic groups through its unique aesthetic symbols. However, in modern society, the protection and inheritance of this traditional craft faces many challenges, such as the aging of inheritors, shrinking market demand, and conflicts with modern aesthetic concepts. This situation makes it difficult for paper cutting to meet the demand for large-scale production. In addition, the lack of research on the effective transformation of traditional culture makes it difficult for paper cutting to meet the fashion needs of contemporary customers, thus limiting its innovation and development space.

[0003] With the rapid popularization of generative artificial intelligence (GenAI), paper-cutting art is undergoing a profound technological empowerment and creative innovation. As a new type of productive force, generative AI not only restructures the creative paradigm but also alleviates the dilemma of inheritance. However, current technology is limited by algorithm maturity and dataset quality, resulting in AI-generated images generally suffering from problems such as feature structure deviations, missing or abnormal details, broken lines, and area adhesion. Furthermore, the high operational threshold of the models means that in precision-sensitive production practices like paper-cutting, the generation process remains highly uncontrollable, making it difficult to directly meet the refined and demanding practical requirements. Currently, mainstream visual detection models can be divided into three categories: traditional object detection models, deep learning-based detection models, and deep detection models for small objects. However, these methods all have limitations when processing paper-cutting images: traditional models have rigid programs, weak semantic understanding, and poor generalization and robustness; while advanced deep learning models, when faced with the complex textures and intricate structures of paper-cutting, also struggle to completely avoid the problems of feature confusion and missed details in small objects.

[0004] Therefore, existing technologies suffer from problems such as uncontrollable quality of AI-generated paper-cutting images, low efficiency in defect detection and screening, easy omission of minor structural defects in small targets (such as tiny breaks or adhesions), and lack of a scientific and systematic evaluation system that combines objective technological limitations with subjective aesthetic and artistic value. Summary of the Invention

[0005] The purpose of this invention is to address the core pain points faced by intangible cultural heritage paper-cutting in digital innovation, such as uncontrollable AI-generated quality, low defect detection efficiency, and lack of a scientific evaluation system. It provides a paper-cutting screening method, equipment, and medium based on target detection and collaborative evaluation. This invention constructs a human-machine collaborative intelligent evaluation system that integrates improved YOLOv11n defect detection and fuzzy comprehensive evaluation (FCE), achieving a closed-loop process from generation, quality inspection, selection to cultural and creative transformation, providing a replicable technical path for the digital inheritance and industrial development of intangible cultural heritage paper-cutting.

[0006] In a first aspect, the present invention provides a paper-cutting screening method based on target detection and collaborative evaluation, comprising the following steps: Step S1: Obtain the original paper-cutting image resources, perform image preprocessing on the original paper-cutting image resources to remove interfering features, and construct a standardized paper-cutting image training dataset. Step S2: Obtain the Flux.1 diffusion model with cross-modal attention mechanism injection, fine-tune the parameters of the Flux.1 diffusion model using the paper-cutting image training dataset, and control the image generation process in the fine-tuned Flux.1 diffusion model by inputting preset theme conditions to generate an initial set of paper-cutting images; Step S3: Obtain the improved YOLOv11n-PPA target detection model. Input the initial set of paper-cut images into the improved YOLOv11n-PPA target detection model for process defect detection and classification statistics. Filter the images based on a preset defect density threshold and output an objective preliminary set of paper-cut images. The improved YOLOv11n-PPA target detection model includes: a backbone network for extracting multi-scale features, a neck network for fusing the multi-scale features, and a detection head for outputting detection results. The backbone network has stacked C3k2_PConvX modules based on partial convolution operators. A global attention mechanism module is embedded before the high-level feature input of the detection head. Step S4: Obtain a multi-dimensional evaluation index set, determine the weight vector of each index in the multi-dimensional evaluation index set based on the expert scoring mechanism, and perform a secondary optimization calculation on the objective preliminary screening paper-cutting image set in combination with the fuzzy comprehensive evaluation model, and output the image with the highest comprehensive score as the optimal paper-cutting image.

[0007] As an optional implementation of the first aspect of this application, in step S2, the Flux.1 diffusion model encodes the images in the paper-cutting image training dataset into the latent space through a pre-trained variational autoencoder; during the training phase, the Flux.1 diffusion model learns a linear interpolation path from the noise distribution to the data distribution through a modified flow model; when generating the initial paper-cutting image set, the Flux.1 diffusion model starts from the initial pure noise and performs iterative denoising processing along the path of solving ordinary differential equations in the learned velocity field, and finally outputs the initial paper-cutting image set.

[0008] As an optional implementation of the first aspect of this application, in step S2, the preset topic condition control includes sketch outline feature conditions and color control text conditions; the cross-modal attention mechanism maps the features from the sketch outline feature conditions to query vectors, and maps the features from the color control text conditions after dual encoder processing to key vectors and value vectors, and realizes the spatial guidance of text on image generation by calculating attention weights; the specific process of fine-tuning the parameters of the Flux.1 diffusion model using the paper-cutting image training dataset is as follows: using low-rank adaptation technology, on the basis of freezing the original model weight matrix of the Flux.1 diffusion model, a low-rank decomposition matrix is ​​introduced to perform forward propagation weight update operation, so that the model learns the paper-cutting pattern features.

[0009] As an optional implementation of the first aspect of this application, in step S3, the backbone network takes the paper-cut image to be detected as the input starting point, and generates multiple scale feature maps with decreasing resolution through downsampling operations in sequence. After each downsampling stage, feature extraction is performed through the stacked C3k2_PConvX modules. The internal processing of the C3k2_PConvX module is as follows: the input features are divided into a first part of channel features and a second part of channel features according to the channel dimension. Spatial convolution operation is performed only on the first part of channel features to extract deep features. The second part of channel features maintains an identity mapping. Then, the first part of channel features after convolution operation and the second part of channel features are concatenated and fused by 1×1 convolution before being output.

[0010] As an optional implementation of the first aspect of this application, the neck network adopts a feature pyramid structure. It concatenates the deep semantic features and shallow detail features output by the backbone network through upsampling and lateral connection operations, and generates an enhanced multi-scale feature map after processing by a module containing partial convolutions. The global attention mechanism module consists of a channel attention module and a spatial attention module connected in series. Its processing of high-level features is as follows: first, the input high-level features are processed by global average pooling, global max pooling, and a multilayer perceptron to obtain channel attention weights. Then, the channel attention weights are multiplied element-wise with the input high-level features to obtain channel attention output features. Subsequently, channel-dimensional pooling and convolution operations are performed on the channel attention output features to obtain spatial attention weights. The spatial attention weights are then multiplied element-wise with the channel attention output features. Finally, after activation and convolution fusion, the output is sent to the detection head.

[0011] As an optional implementation of the first aspect of this application, the process defects include red line defects, white dot defects, and white line defects; the improved YOLOv11n-PPA target detection model outputs predicted bounding boxes, defect categories, and confidence scores, and the loss function used for network parameter optimization consists of three weighted parts: a binary cross-entropy classification loss for evaluating the difference between the predicted category and the true category, an intersection-union ratio (IUU) localization loss for evaluating the overlap between the predicted bounding box and the true bounding box, and a binary cross-entropy confidence score loss for evaluating the accuracy of the target existence probability prediction.

[0012] As an optional implementation of the first aspect of this application, in step S4, the multi-dimensional evaluation index set includes process feasibility index, aesthetic innovation index, cultural heritage index, market potential index, and artistic value index, and configures corresponding comment levels and level assignment vectors for the multi-dimensional evaluation index set; the process of determining the weight vector of each index based on the expert scoring mechanism includes: obtaining quantitative scoring data of the importance of each index from multiple experts, summing all expert scores for the same index, dividing by the total expert scores of all indicators for normalization, and outputting the weight components of each index to form the weight vector, wherein the weight component corresponding to the process feasibility index has the highest value.

[0013] As an optional implementation of the first aspect of this application, the weight vectors of each indicator in the multi-dimensional evaluation index set are determined based on an expert scoring mechanism. A secondary optimization calculation is then performed on the objective preliminary screening set of paper-cutting images using a fuzzy comprehensive evaluation model. The image with the highest comprehensive score is output as the optimal paper-cutting image. The specific process includes: for the images to be evaluated in the objective preliminary screening set, statistical analysis of the voting data of multiple evaluators on each evaluation index and each comment level is conducted, and the proportion of each level is calculated to construct the fuzzy evaluation matrix corresponding to the image to be evaluated; a weighted average operator is used to perform matrix multiplication between the weight vector and the fuzzy evaluation matrix to obtain the comprehensive evaluation vector corresponding to the image to be evaluated; the comprehensive evaluation vector is multiplied by the level assignment vector to calculate the comprehensive score of the image to be evaluated; and the comprehensive scores of all images to be evaluated are sorted in descending order, and the image at the top of the sorted sequence is output as the final optimal paper-cutting image.

[0014] In a second aspect, embodiments of this application provide an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the method described in the first aspect.

[0015] Thirdly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The quality and controllability of generative paper-cutting data are improved by dual condition constraints and parameter fine-tuning. Addressing the issues of scarce paper-cutting image data and inconsistent generation quality, this invention introduces dual cross-modal condition control of sketch and color in the Flux.1 diffusion model, and combines this with LoRA (low-rank adaptation) for efficient parameter fine-tuning. This technical feature effectively suppresses random structural mutations during the generation process, enabling AI to accurately learn the unique artistic rules of paper-cutting patterns, significantly improving the cultural fidelity and process standardization of paper-cutting generation, and providing a high-quality, standardized data source for subsequent screening.

[0017] (2) The innovative and improved YOLOv11n-PPA network significantly improves the efficiency and accuracy of target detection of minute process defects. In view of the problem that traditional detection networks have a large amount of computation and are prone to missed detection when processing complex hollow textures, this invention introduces a C3k2_PConvX module based on partial convolution operators into the backbone network, which greatly reduces computational redundancy and improves detection speed. At the same time, a global attention mechanism (GAM) module is embedded in the detection head, which effectively enhances the network's ability to focus on global context features and local fine defects such as broken lines and waste residue, and realizes efficient and accurate screening of objective process defects.

[0018] (3) A closed-loop scientific evaluation system that integrates objective algorithm quality inspection with subjective human aesthetics has been constructed. This invention breaks through the technical bottleneck of traditional generative AI's "blind image generation and blind human selection," creatively using YOLO objective defect detection as a prerequisite for rigorous process screening, followed by the introduction of a fuzzy comprehensive evaluation (FCE) model for secondary optimization, with "process feasibility" given the highest weight. This technical solution achieves seamless integration of bottom-level data generation, intermediate-level machine vision quality inspection, and top-level multi-dimensional (culture, aesthetics, market) human-machine collaborative evaluation, providing an extremely efficient and replicable technical paradigm for the large-scale industrialization of intangible cultural heritage skills and the development of cultural and creative products. Attached Figure Description

[0019] Figure 1 This is a flowchart of the implementation method provided in the embodiments of the present invention; Figure 2 This is a diagram of the improved YOLOv11n-PPA network architecture in an embodiment of the present invention; Figure 3 This is a flowchart of the Global Attention Mechanism (GAM) module in an embodiment of the present invention; Figure 4 These are some illustrations of the paper-cutting materials collected and pre-processed in this embodiment of the invention; Figure 5 This is a partial illustration of a high-quality Chinese zodiac horse paper-cut image generated by the Flux.1 diffusion model in an embodiment of the present invention; Figure 6 This is a training process curve of the YOLOv11n-PPA model in an embodiment of the present invention; Figure 7 This is a graph showing the precision, recall, and F1 score of the YOLOv11n-PPA model in paper-cutting defect detection in this embodiment of the invention. Figure 8 Examples of high-quality paper-cutting images obtained after objective initial screening in embodiments of the present invention; Figure 9 This is an example of the optimal paper-cutting image selected through fuzzy comprehensive evaluation in this embodiment of the invention; Figure 10 This is a rendering of a cultural and creative application derived from an optimal paper-cutting image in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] Example In this embodiment, by generating high-quality zodiac horse festival paper-cutting samples, a closed-loop process from generation, quality inspection, selection to cultural and creative transformation is achieved, providing a replicable technical path for the digital inheritance and industrial development of intangible cultural heritage paper-cutting. Please refer to [link / reference]. Figure 1 This invention provides an implementation process for a paper-cutting screening method based on target detection and collaborative evaluation. To achieve the above objectives, the invention employs the following implementation steps: Step S1: Obtain the original paper-cutting image resources, perform image preprocessing on the original paper-cutting image resources to remove interfering features, and construct a standardized paper-cutting image training dataset.

[0023] Specifically, the details of step S1 in creating the exclusive zodiac horse paper-cutting material are as follows: To generate innovative paper-cutting works using the Flux.1 model, the model first needed to fully learn existing paper-cutting art styles. Therefore, in the first stage, approximately 415 paper-cutting images were successfully collected from libraries, paper-cutting art books, and multiple art websites. These original images were then systematically preprocessed, including removing background interference, standardizing image sizes, and correcting shooting angles to ensure the standardization and consistency of the input data. Furthermore, each image was labeled with the work's title and author information for easy retrieval and traceability later. After this standardization process, a paper-cutting resource library containing the training dataset of paper-cutting images was finally constructed, laying a solid foundation for model training.

[0024] Step S2: Obtain the Flux.1 diffusion model with cross-modal attention mechanism injection, fine-tune the parameters of the Flux.1 diffusion model using the paper-cutting image training dataset, and control the image generation process in the fine-tuned Flux.1 diffusion model by inputting preset topic conditions to generate an initial set of paper-cutting images.

[0025] Specifically, based on the Flux.1 diffusion model, a series of 500 paper-cut images of the Chinese zodiac horse with five festival themes are generated. The specific implementation method of step S2 is as follows: Step S21, Training Phase First, the paper-cut image is encoded into the latent space using a pre-trained VAE. During the training phase, the model learns a linear interpolation path from the noise distribution to the data distribution, with the optimization objective being to minimize the velocity field prediction error. The model then learns to predict the velocity field. That is, the vector direction from noise to data: In the formula That is, the actual velocity direction. express, This represents the initial latent variable (pure noise). Represents the latent variables of the target data. This indicates the time step. Compared to traditional diffusion models that predict noise, the RectifiedFlow model achieves faster sampling speeds and a more stable training process through deterministic paths.

[0026] Step S22, Ordinary Differential Equations (ODE) When the initial paper-cut image is generated, it starts from pure noise. Starting from the learned velocity field, solve the ordinary differential equations: in Represents a d-dimensional identity matrix. Let be the latent variable at time t. ( () is a parameterized velocity field model.

[0027] Step S23, injection of cross-modal attention mechanism Simultaneously, graph contour feature conditions and color-controlled text conditions are introduced. These dual topic conditions are injected into the Transformer backbone of the Flux.1 diffusion model through a cross-modal attention mechanism. The calculation method is as follows: Among them, query Feature mapping derived from sketch contour feature conditions, key Sum The feature map derived from the color-controlled text conditions, after being processed by a dual encoder, uses attention weights to spatially guide the text in image generation.

[0028] Step S24, fine-tune Lora To adapt to the style of paper-cutting for the Chinese zodiac horse, LoRA was used to efficiently fine-tune the parameters of the Flux.1 diffusion model to generate an initial set of paper-cutting images. The principle is as shown in equation (4), which effectively learns the artistic rules of paper-cutting patterns while retaining the model's general generation capabilities. The core principle of LoRA is to freeze the weights of the original model. Based on this, a low-rank decomposition matrix is ​​introduced. and This makes forward propagation become: in This represents the model's input feature vector; This represents the output features after weight updates; This represents the frozen weight matrix of the original model; , This represents a low-rank decomposition matrix, used for efficient parameter fine-tuning.

[0029] In this way, the model can effectively learn the artistic rules of paper-cutting patterns while retaining its general generative capabilities.

[0030] Step S3: Obtain the improved YOLOv11n-PPA target detection model. Input the initial set of paper-cut images into the improved YOLOv11n-PPA target detection model for process defect detection and classification statistics. Filter the images based on a preset defect density threshold and output an objective preliminary set of paper-cut images. The improved YOLOv11n-PPA target detection model includes: a backbone network for extracting multi-scale features, a neck network for fusing the multi-scale features, and a detection head for outputting detection results. The backbone network has stacked C3k2_PConvX modules based on partial convolution operators. A global attention mechanism module is embedded before the high-level feature input of the detection head.

[0031] Specifically, step S3 involves using an improved YOLOv11n-PPA model to detect and classify process defects in the generated paper-cutting. These defects include red line defects, white spot defects, and white line defects. The specific implementation method for filtering based on a preset defect density threshold is as follows: like Figure 2As shown, the improved YOLOv11n-PPA object detection model consists of three parts: a backbone network, a neck network, and a head. The backbone network achieves efficient multi-scale feature extraction through stacked convolutional layers and the C3k2_PConvX module. The neck network employs a feature pyramid structure to fuse deep semantic features with shallow detail features. The head embeds a global attention mechanism (GAM) module after high-level features, such as... Figure 3 As shown.

[0032] Step S31, embedding the C3k2_PConvX module First, Backbone uses the paper-cut image to be detected. Starting with H, where H represents the image height and W represents the image width, the image is sequentially downsampled through convolutions with a stride of 2, generating feature maps P3 (1 / 8), P4 (1 / 16), and P5 (1 / 32) with decreasing resolution. After each downsampling stage, two C3k2_PConvX modules are stacked for feature extraction, with the number of channels being 256, 512, and 1024 respectively.

[0033] The C3k2_PConvX module is a lightweight design based on the CSP architecture, with its core operator being Partial Convolution (PConv). PConv performs convolution only on a subset of the input channels, maintaining the identity mapping for the remaining channels to reduce computational cost. Let the input features... Divided by channel ,in (generally C represents the total number of channels in the input feature map, then PConv represents the partial convolution output features, and the output is: in, A 3×3 convolution is used for partial convolution. In C3k2_PConvX, the input is first compressed by a 1×1 convolution, and then split into two branches: one branch extracts deep features through multiple bottleneck layers containing PConv, and the other branch is directly connected; finally, the two branches are concatenated and fused by a 1×1 convolution for output.

[0034] Step S32, Neck Feature Pyramid Enhancement Secondly, the Neck module employs a feature pyramid structure, fusing multi-scale features through upsampling and lateral connections. Specifically, it starts by upsampling from P5, the output of the backbone network, concatenating it with P4, and then passing it through the C3k2 module to obtain a new P4; it is then upsampled again and concatenated with P3, passing it through the C3k2 module to obtain a new P3. Subsequently, the new P3 and P4 are downsampled, concatenated with the corresponding layers, and then passed through the C3k2 module again to generate the enhanced features P4 and P5.

[0035] Step S33, Insert the GAM module Finally, the head performs object detection on feature maps at three scales: P3, P4, and P5. Specifically, a GAM module is inserted after feature map P5 to strengthen global dependencies. GAM consists of concatenated channel attention and spatial attention: given high-level input features… Channel attention: in, Sigmoid is a dedicated activation function; This represents a multilayer perceptron; Indicates global average pooling; Indicates global max pooling; This represents the channel attention weight.

[0036] Spatial attention: in, 7×7 convolutions are used specifically for spatial attention. This indicates average pooling along the channel dimension. This indicates max pooling along the channel dimension; This represents the spatial attention weights.

[0037] Fusion Output: in Indicates the input feature map; This represents the output feature map after channel and spatial attention fusion; This is element-wise multiplication.

[0038] The enhanced P5, along with P3 and P4, is fed into the detection head layer, outputting the predicted bounding box, defect category, and confidence score. The loss function adopts the standard YOLO series form: in For binary cross-entropy classification loss, Locate the loss for CIoU. For binary cross-entropy confidence loss, , , These are the weighting coefficients for the corresponding loss function, used to balance the various loss terms.

[0039] Step S4: Obtain a multi-dimensional evaluation index set, determine the weight vector of each index in the multi-dimensional evaluation index set based on the expert scoring mechanism, and perform a secondary optimization calculation on the objective preliminary screening paper-cutting image set in combination with the fuzzy comprehensive evaluation model, and output the image with the highest comprehensive score as the optimal paper-cutting image.

[0040] Specifically, the implementation method of step S4, which uses the FCE evaluation method to perform secondary optimization on the objectively screened paper-cutting images based on objective screening, is as follows: Step S41, Establishment of the evaluation index system The paper-cutting images were evaluated from five dimensions: aesthetic innovation, cultural heritage, artistic value, technological feasibility, and market potential. (This is a set of factors.) Collection of comments Corresponding to the 5 levels {Excellent, Good, Average, Poor, Very Poor}, the score vector for each level is as follows: Among them, process feasibility, as an indicator directly related to the initial defect detection, has the highest weight in the subsequent weight determination.

[0041] Step S42, Determine the indicator weights Five experts in the field of paper-cutting were invited to independently assign importance values ​​to each indicator using a percentage-based system. Let the [indicator name be missing]. Experts on the indicators The rating is ( ), then the indicator The weights are: Obtain the weight vector ,satisfy .

[0042] Step S43, Construction of fuzzy evaluation matrix For the 25 images to be evaluated, 50 evaluators were effectively invited to rate each image based on five indicators. For the... Image (of which) Image sequence number =1,…,25), statistical indicators In the comments The number of votes received Calculate membership degree The fuzzy evaluation matrix is ​​obtained as follows: Step S44, Calculation of overall score Weight vectors are synthesized using a weighted average operator. With fuzzy matrix , obtained the Comprehensive evaluation vector of images Combining comment level scoring vector Calculate the overall score of the image. in This represents the overall membership degree of the j-th rating for the r-th image; This represents the score assigned to the j-th comment level; This represents the overall evaluation score of the r-th image.

[0043] The 25 images were sorted from highest to lowest score, and the one with the highest score was the best paper-cutting image.

[0044] Example verification The following example, using a festive series of zodiac horse paper-cuts, further illustrates the method of this invention.

[0045] Step S10: Collect materials for the Chinese zodiac horse paper-cutting; To generate innovative paper-cutting works using the Flux.1 model, the model first needed to fully learn existing paper-cutting art styles. Therefore, in the first stage, approximately 415 paper-cutting images were successfully collected through libraries, paper-cutting art books, and multiple art websites. These images were then systematically preprocessed, including removing background interference, standardizing image sizes, and correcting shooting angles to ensure the standardization and consistency of the input data. Furthermore, each image was labeled with the work's title and author information for easy retrieval and traceability later. After this standardization process, a high-quality paper-cutting resource library was finally constructed (e.g., ...). Figure 4 As shown in the figure, this lays a solid foundation for model training.

[0046] Step S20: The Flux.1 diffusion model generates a paper-cut image of the Chinese zodiac horse. After constructing the paper-cutting material library, we used it as a training dataset and input it into the Flux.1 diffusion model. We then fine-tuned the model using LoRA technology to enhance its ability to express the style of paper-cutting art. To verify the model's generation effect in specific cultural contexts, we selected five major traditional Chinese festivals—Spring Festival, Lantern Festival, Qixi Festival, Dragon Boat Festival, and Mid-Autumn Festival—as creative themes and conducted image generation experiments using the trained LoRA model. During the generation process, we not only input the festival-related sketch outlines and color control information but also used LLM to refine and expand the keywords for each theme, thereby improving the expressiveness and diversity of the prompts. Finally, we generated over a hundred paper-cutting images for each theme. After objective initial screening and designer selection to remove samples that did not meet expectations, we retained 60 high-quality zodiac horse paper-cutting works for each theme, such as... Figure 5 As shown.

[0047] Step S30, YOLOv11n-PPA objective preliminary assessment of paper-cutting defects; To ensure the objectivity and professionalism of the data, this invention interviewed artisans with more than five years of experience in the paper-cutting industry. Based on their provided process standards, a quantitative basis for defining defects was established: the width of a single thin line should not be less than 0.5 mm, and the width of a hollowed-out portion should not be less than 1 mm. Based on this, we defined three defect types in the data annotation: red lines, white dots, and white lines. All annotation work was completed in Label Studio, as shown. To enhance the model's generalization ability, we performed data augmentation operations on the original data, including local defect cropping, rotation, and brightness adjustment. The dataset was ultimately expanded to 1000 images and divided into training, validation, and test sets in a 5:2:3 ratio, containing 500, 200, and 300 images respectively. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and model selection during the training process, and the test set consists of 300 uncropped paper-cut images (60 for each of the five themes) generated by the Flux.1 diffusion model to evaluate the model's ability to generalize defect detection on the generated images.

[0048] To evaluate the performance of the YOLOv11n-PPA model in defect recognition and detection of paper-cut images, we selected precision, recall, and mean precision (mAP) as evaluation metrics. In the definition of mAP, "m" represents the average value, and AP@0.5 represents the average precision of samples in that class when the intersection-union ratio (IUU) threshold in the confusion matrix is ​​0.5. mAP@0.5 is the average precision value across all classes, reflecting the trend of model precision changing with recall. A higher value indicates a stronger ability of the model to maintain high precision even under high recall conditions.

[0049] In the analysis of experimental results, we first used the training process curve (such as...) Figure 6 The convergence and performance of the model were evaluated (as shown). In terms of detection performance, as the number of training epochs increased, the bounding box loss, clipping loss, and diff loss all showed a decreasing trend and gradually converged. Meanwhile, precision and recall steadily improved, indicating that the model was sufficiently trained and did not exhibit significant overfitting. These results demonstrate that the constructed YOLOv11n-PPA model can efficiently identify process defects such as broken lines and overlapping cutouts, providing a reliable initial screening basis for the subjective aesthetic evaluation of subsequent AI-generated paper-cutting images. This verifies the effectiveness of this method in the digital preservation of intangible cultural heritage paper-cutting.

[0050] To further quantify the model's performance in the AI-generated paper-cutting defect detection task, Figure 7 The corresponding evaluation curves are shown in the figure. As can be seen from the figure, the overall F1 score peaks at 0.81 (corresponding to a confidence level of 0.327), the precision reaches 1.00 at a confidence level of 0.903, and the mAP@0.5 is 0.869. The mean precision (AP) for the three defect types (red_line, white_dot, white_line) are 0.878, 0.857, and 0.874, respectively. The PR curves generally approach the ideal state, and both the F1 and recall curves exhibit good confidence-to-priority trade-offs, indicating that the model maintains high recall while also possessing excellent precision, demonstrating balanced detection capability and generalization for different types of paper-cutting defects.

[0051] Given the good detection performance of the model in the previous stage of research, we further applied it to 300 test set images that were not used in the training phase, and used this data to select samples for subsequent analysis. Specifically, the trained model was used to detect defects in each image in the validation set, and the number and category distribution of defects in each image were statistically analyzed. A comprehensive statistical analysis of the images is shown in Table 1 below. For ease of reference, A represents the Spring Festival; B represents the Lantern Festival; C represents the Mid-Autumn Festival; D represents the Qixi Festival; and E represents the Dragon Boat Festival. Subsequently, the 5 images with the fewest defects were selected from each theme category. If the number of images meeting the minimum defect criterion for a particular theme exceeded 5, a focus group composed of design experts with more than five years of experience was invited to discuss and rank these candidate images, ultimately determining the top 5 images for that theme. Through this process, a total of 25 samples were obtained, as shown below. Figure 8 As shown, it is used for subsequent fuzzy comprehensive evaluation.

[0052] Table 1. Statistics on the number of defects detected by YOLOv11n-PPA target detection Step S40, Fuzzy comprehensive evaluation and ranking In the previous stage, after objective initial screening by YOLOv11n-PPA and careful selection by the design expert group, a total of 25 paper-cutting images that meet the basic manufacturing process requirements were selected. Figure 8 As shown in Table 2, five experts in the field of paper-cutting were invited to independently assign values ​​to the importance of each indicator using a percentage system. The scoring results of each expert are shown in Table 2, where the feasibility of the process is represented by PF; aesthetic innovation by AI; cultural heritage by CT; market potential by MP; and artistic value by AV. The weight vector is calculated according to formula (10). The corresponding order of each component is: technological feasibility (indicator 1), aesthetic innovation (indicator 2), cultural heritage (indicator 3), market potential (indicator 4), and artistic value (indicator 5).

[0053] Table 2. Determination of Indicator Weights To ensure more accurate results, a quantitative survey method was used to distribute questionnaires. The 25 paper-cutting images were scored on five indicators, with scores ranging from 1 to 5. To ensure the validity of the questionnaires, they were completed by design students, paper-cutting enthusiasts, and the general public. A total of 63 questionnaires were distributed. After data cleaning, 13 samples with a completion time of less than 3 minutes, identical options, and obvious patterns in the answers were removed, resulting in 50 valid questionnaires. Experts can give comments and scores based on the corresponding "very good, good, average, poor, very poor" for each image. The membership degree is calculated according to formula (11) to obtain the results, as shown in Table 3.

[0054] Table 3: Membership Matrix of Each Image Indicator (Each row represents an image, each of the five indicators occupies a column, and each column contains a membership vector of five levels) Based on equation (12-13), the complete fuzzy comprehensive evaluation results of 25 paper-cutting images are obtained and sorted from high to low according to the comprehensive score. The comprehensive evaluation vector (five dimensions, in the order of "very good, good, average, poor, very poor") and comprehensive score of each image are also listed (Table 4).

[0055] Table 4: Fuzzy Comprehensive Evaluation Results of 25 Paper-cut Images (Sorted by Score in Descending Order) The image with the highest overall score from each of the five themes was selected as the representative work for that theme. Figure 9As shown in the figure. The results show that: in the Spring Festival theme, A5 scored the highest (4.0544); in the Lantern Festival theme, B2 scored the highest (4.3016); in the Mid-Autumn Festival theme, C5 scored the highest (4.2032); in the Qixi Festival theme, D1 scored the highest (3.8448); and in the Dragon Boat Festival theme, E1 scored the highest (3.8288).

[0056] In summary, this invention, based on the objective elimination of images with fundamental technological defects using the YOLOv11n-PPA model and combined with the fuzzy comprehensive evaluation method, ultimately selected five superior paper-cutting images. The selected images excel in five dimensions: technological feasibility, aesthetic innovation, cultural heritage, market potential, and artistic value. They possess both highly recognizable festival cultural symbols and clear, convertible graphic structures. Furthermore, zodiac elements have a broad public recognition base and stable market demand, particularly fitting the consumption scenarios of festival gifts and zodiac-themed cultural and creative products, effectively enhancing the product's market recognition and commercial value. A simulated effect image of the selected best zodiac horse paper-cutting was created using Photoshop, as shown below. Figure 10 As shown, transforming and applying it in the form of cultural and creative products also injects a more contemporary way of expression into traditional paper-cutting art.

[0057] Optionally, embodiments of this application also provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of a paper-cutting screening method based on target detection and collaborative evaluation, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0058] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of a paper-cutting screening method based on target detection and collaborative evaluation, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0059] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0060] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0062] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A paper-cutting screening method based on target detection and collaborative evaluation, characterized in that, Includes the following steps: Step S1: Obtain the original paper-cutting image resources, perform image preprocessing on the original paper-cutting image resources to remove interfering features, and construct a standardized paper-cutting image training dataset. Step S2: Obtain the Flux.1 diffusion model with cross-modal attention mechanism injection, fine-tune the parameters of the Flux.1 diffusion model using the paper-cutting image training dataset, and control the image generation process in the fine-tuned Flux.1 diffusion model by inputting preset theme conditions to generate an initial set of paper-cutting images; The Flux.1 diffusion model encodes the images in the paper-cutting image training dataset into the latent space through a pre-trained variational autoencoder. During the training phase, the Flux.1 diffusion model learns a linear interpolation path from the noise distribution to the data distribution through a modified flow model. When generating the initial paper-cutting image set, the Flux.1 diffusion model starts from the initial pure noise and performs iterative denoising processing along the path of solving ordinary differential equations in the learned velocity field, finally outputting the initial paper-cutting image set. Step S3: Obtain the improved YOLOv11n-PPA target detection model, input the initial paper-cutting image set into the improved YOLOv11n-PPA target detection model for process defect detection and classification statistics, filter based on the preset defect density threshold, and output an objective preliminary screening paper-cutting image set. The process defects include red line defects, white dot defects, and white line defects; the improved YOLOv11n-PPA target detection model outputs predicted bounding boxes, defect categories, and confidence scores. The loss function used for network parameter optimization consists of three weighted parts: a binary cross-entropy classification loss used to evaluate the difference between the predicted category and the true category, an intersection-union ratio localization loss used to evaluate the overlap between the predicted bounding box and the true bounding box, and a binary cross-entropy confidence score loss used to evaluate the accuracy of the target existence probability prediction. The improved YOLOv11n-PPA target detection model includes: a backbone network for extracting multi-scale features, a neck network for fusing the multi-scale features, and a detection head for outputting detection results. The backbone network has stacked C3k2_PConvX modules based on partial convolution operators, and a global attention mechanism module is embedded before the high-level feature input of the detection head. The backbone network takes the paper-cut image to be detected as the input starting point, and generates multiple scale feature maps with decreasing resolution through downsampling operations in sequence. After each downsampling stage, feature extraction is performed through the stacked C3k2_PConvX modules. The internal processing of the C3k2_PConvX module is as follows: the input features are divided into a first part of channel features and a second part of channel features according to the channel dimension. Spatial convolution operation is performed only on the first part of channel features to extract deep features. The second part of channel features maintains an identity mapping. Then, the first part of channel features after convolution operation and the second part of channel features are concatenated and output after 1×1 convolution fusion. Step S4: Obtain a multi-dimensional evaluation index set, determine the weight vector of each index in the multi-dimensional evaluation index set based on the expert scoring mechanism, and perform a secondary optimization calculation on the objective preliminary screening paper-cutting image set in combination with the fuzzy comprehensive evaluation model, and output the image with the highest comprehensive score as the optimal paper-cutting image.

2. The method according to claim 1, characterized in that, In step S2, the preset theme condition control includes sketch outline feature conditions and color control text conditions; The cross-modal attention mechanism maps the features from the sketch contour feature conditions to query vectors, and maps the features from the color control text conditions after dual encoder processing to key vectors and value vectors. By calculating attention weights, it realizes the spatial guidance of text on image generation. The specific process of fine-tuning the parameters of the Flux.1 diffusion model using the paper-cutting image training dataset is as follows: using low-rank adaptation technology, on the basis of freezing the original model weight matrix of the Flux.1 diffusion model, a low-rank decomposition matrix is ​​introduced to perform a forward propagation weight update operation, so that the model learns the paper-cutting pattern features.

3. The method according to claim 1, characterized in that, The neck network adopts a feature pyramid structure. It concatenates the deep semantic features and shallow detail features output by the backbone network through upsampling and lateral connection operations. After processing by a module containing partial convolution, it generates an enhanced multi-scale feature map. The global attention mechanism module consists of a channel attention module and a spatial attention module connected in series. The process of processing high-level features is as follows: First, the input high-level features are processed by global average pooling, global max pooling, and multilayer perceptron to obtain channel attention weights. Then, the channel attention weights are multiplied element-wise with the input high-level features to obtain channel attention output features. Subsequently, the channel attention output features are processed by channel dimension pooling and convolution to obtain spatial attention weights. Then, the spatial attention weights are multiplied element-wise with the channel attention output features. Finally, after activation and convolution fusion, the output is sent to the detection head.

4. The method according to claim 1, characterized in that, In step S4, the multi-dimensional evaluation index set includes process feasibility index, aesthetic innovation index, cultural heritage index, market potential index and artistic value index, and configures corresponding evaluation level and level scoring vector for the multi-dimensional evaluation index set. The process of determining the weight vector of each indicator based on the expert scoring mechanism includes: obtaining quantitative scoring data of the importance of each indicator from multiple experts, summing all expert scores for the same indicator, dividing by the total expert scores of all indicators for normalization, and outputting the weight components of each indicator to form the weight vector, wherein the weight component corresponding to the process feasibility indicator has the highest value.

5. The method according to claim 4, characterized in that, The weight vectors of each indicator in the multi-dimensional evaluation index set are determined based on an expert scoring mechanism. A secondary optimization calculation is then performed on the objectively screened paper-cutting image set using a fuzzy comprehensive evaluation model. The image with the highest comprehensive score is output as the optimal paper-cutting image. The specific process includes: For the images to be evaluated in the objective preliminary screening set of paper-cutting images, the voting data of multiple evaluators on the image to be evaluated in various evaluation indicators and various comment levels are statistically analyzed, and the proportion of each level is calculated to construct the fuzzy evaluation matrix corresponding to the image to be evaluated. A weighted average operator is used to perform matrix multiplication between the weight vector and the fuzzy evaluation matrix to obtain the comprehensive evaluation vector corresponding to the image to be evaluated. The comprehensive evaluation vector and the grade assignment vector are multiplied by a dot product to calculate the comprehensive score of the image to be evaluated. The images are sorted in descending order based on their overall scores, and the image at the top of the sorted sequence is output as the final optimal paper-cutting image.

6. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of a paper-cutting screening method based on target detection and collaborative evaluation as described in any one of claims 1-5.

7. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the steps of the paper-cutting screening method based on target detection and collaborative evaluation as described in any one of claims 1-5.

Images