Denoising preprocessing-based pattern extraction method

Abstract

The invention discloses a pattern extraction method based on denoising preprocessing, and the method comprises the following steps: 1, carrying out model training and optimization, carrying out data preparation, inputting image data, and dividing a data set into a training set, a test set and a verification set; step 2, extracting and storing a pattern; performing de-noising processing on a to-be-extracted pattern image by using a non-restrictive median filtering technology, then performing pattern extraction by using a trained neural network model based on deep perception fusion, and converting the pattern into a final storable vector diagram format through an image processing technology; compared with the prior art, the method has the advantages that the pattern information in the image can be accurately extracted, a high-precision scalable vector diagram format is generated and stored, and the edge features of the intangible cultural heritage pattern are extracted and stored through denoising preprocessing and a deep learning algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and specifically, to a pattern extraction method based on denoising preprocessing. Background Art

[0002] With the acceleration of the globalization process, Intangible Cultural Heritage (ICH) is facing unprecedented challenges. Especially in the digital preservation of artistic patterns and traditional handicrafts, how to effectively protect and inherit these precious cultural heritages by using modern technologies has become an urgent problem to be solved.

[0003] The digitalization of traditional handicraft patterns, especially complex patterns, often faces image quality problems, such as noise interference, detail loss, and image edge blurring. Specifically, these problems are particularly prominent in the image acquisition stage. Especially under the condition of high noise, traditional image processing methods cannot accurately retain the pattern details, which affects the accuracy of subsequent pattern extraction and vectorization.

[0004] Taking Qiang embroidery as an example, as a traditional craft of an ancient ethnic minority in China, its patterns are not only rich in cultural characteristics but also contain rich symbolic meanings. However, the existing digital preservation methods still have technical deficiencies. Especially in the image processing stage, the influence of noise on the pattern edge is relatively large, resulting in the difficulty of accurately restoring the pattern details by traditional edge extraction technologies, which affects the subsequent vectorization processing and digital archiving.

[0005] The existing technologies mainly face the following problems:

[0006] 1. Image noise interference: Traditional image processing methods are difficult to effectively remove noise. Especially in the case of complex patterns, noise may obscure important details, resulting in inaccurate edge extraction.

[0007] 2. Inaccurate edge extraction: When the existing edge extraction algorithms process complex patterns, they are prone to losing details, especially small edges or the structures of complex patterns. This makes it difficult for traditional methods to accurately restore the details of the patterns.

[0008] 3. Poor pattern vectorization effect: Due to inaccurate edge extraction, the subsequent pattern vectorization processing has a poor effect and cannot generate high-quality Scalable Vector Graphics (SVG), which causes troubles for the digital archiving and inheritance of intangible cultural heritages.

[0009] The information disclosed in this background art section is only intended to enhance the overall understanding of the present invention and should not be regarded as an admission or any form of suggestion that this information constitutes the prior art already known to those of ordinary skill in the art. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to overcome the above technical defects, and provide a pattern extraction method based on denoising preprocessing. Through the steps of denoising, edge extraction, pattern recognition, and vectorization processing, the pattern information in the image is accurately extracted, and a high-precision scalable vector graphic format is generated for storage. The edge features of intangible cultural heritage patterns are extracted and saved through denoising preprocessing and deep learning algorithms.

[0011] To solve the above problems, the technical solution of the present invention is a pattern extraction method based on denoising preprocessing, including the following steps:

[0012] Step 1: Model training and optimization;

[0013] First, data preparation is carried out. Input image data, and divide the data set into a training set, a test set, and a validation set. The training set is used for model training, the test set is used to verify the performance of the model, and the validation set is used to evaluate the performance of the model on new data and tune hyperparameters;

[0014] The core of the model is based on a neural network for depth perception fusion. The network structure includes: a perception fusion sub-network and a depth sampling sub-network; the model uses the provided original image and its corresponding paired edge image for integration and optimization, and tests and verifies the model through the test set and the validation set to obtain an optimized neural network model based on depth perception fusion and model parameters;

[0015] Step 2: Pattern extraction and storage;

[0016] First, use non-restrictive median filtering technology to denoise the pattern image to be extracted, and then use the trained neural network model based on depth perception fusion to extract the pattern, and convert it into a final vector graphic format that can be saved through image processing technology. The pattern extraction and storage include image acquisition and digitization, denoising preprocessing, edge extraction, and vectorization processing and storage output.

[0017] Further, in step 1, the input image data includes each original image and its corresponding paired edge image. The original image is used to provide detailed information of the pattern, and the edge image is used to provide edge features as the expected output.

[0018] Further, in step 1, the perception fusion sub-network includes five output blocks similar to encoders, and the output blocks are K-1 to K-5. Each output block is composed of a combination of multiple smaller sub-blocks;

[0019] Among them, K - 1 blocks each have one sub - block. The sub - block has a convolutional layer s2 with a stride of 2, a convolutional kernel size of 3x3, 32 convolutional kernels, and a convolutional layer with 64 convolutional kernels, a convolutional kernel size of 3x3, performing a regular convolution operation without changing the stride; K - 2 blocks have two sub - blocks. One sub - block has two convolutional layers with 128 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2. The other sub - block is a max - pooling layer, a 3x3 max - pooling layer with a stride of 2; K - 3 has three sub - blocks. Two of the sub - blocks each have two convolutional layers with 256 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2. The other sub - block is a max - pooling layer; K - 4 has four sub - blocks. Three of the sub - blocks each have two convolutional layers with 512 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2. The other sub - block is a max - pooling layer; K - 5 has three sub - blocks, and each of the three sub - blocks has two convolutional layers with 512 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2;

[0020] The feature maps generated by each output block are input into an independent depth sampling sub - network. Before merging with the main connection, an average operation of the edge connection is performed to generate an intermediate edge pattern map. Finally, these intermediate edge maps are fused and spliced together to form a learning stack. The last step of the network fuses these features to generate the final pattern map.

[0021] Furthermore, the depth sampling sub - network includes two output blocks, namely b - 1 and b - 2. The output blocks process the feature maps through a conditional stacking structure. Each block contains a convolutional layer and a transposed convolutional layer. The depth sampling sub - network is responsible for up - sampling the extracted intermediate edge maps and finally generating edge maps with the same size as the original image;

[0022] b - 1 processes the input using a 1×1 convolutional kernel, then passes through the ReLU activation function, and then performs a transposed s×s convolution, where the convolutional kernel size s is determined by the scale level of the input feature map; b - 2 is only activated when the input feature map needs to be scaled from the separable sub - network. This block processes iteratively until the size of the feature map reaches twice the target size. When the condition is met, the feature map is fed back to Block - 1.

[0023] Furthermore, in step two, the steps of image acquisition and digitization include: according to different application requirements, the image is obtained by shooting a physical image with a professional photographic device or by designing a pattern on an online platform to obtain a digital image; the collected image data has high definition and high resolution to ensure that the subsequent processing can accurately extract the details of the pattern; the collected image is converted into a standard digital image format for subsequent processing.

[0024] Further, the steps of the noise reduction preprocessing include: performing grayscale processing on the acquired image, calculating the weighted average of the red, green, and blue color channels of each pixel to obtain its grayscale value. Impurities in the image during the shooting process increase the noise of the image, and a non-restrictive median filtering denoising technique is used.

[0025] Further, the non-restrictive median filtering algorithm includes the following steps:

[0026] S1. Set an initial window size, initially set to 3x3. The window will slide pixel by pixel in the image, and calculate the median of each pixel.

[0027] S2. During the filtering process, the algorithm does not restrict the specific size of the window. It decides whether to expand or shrink the window by checking the pixel value changes within the current window. If the noise within the window is strong, the algorithm will increase the size of the window to better smooth the noise. Each time the window size is expanded, it can be expanded up to a fixed upper limit, set to 10x10.

[0028] S3. For each pixel point, calculate the median of all pixel values in its window. The median refers to the middle value after sorting the pixel values in the window. In some cases, if the pixel value within the window differs significantly from the value of the central pixel, then this pixel is considered noise and is replaced with the median.

[0029] S4. Determine whether this pixel is noise based on the pixel value changes within the current window. If the current pixel has a significant difference from the pixels in its neighborhood, then this pixel is considered noise and is replaced with the median. The judgment criterion is through the following steps:

[0030] Calculate the maximum value Max and the minimum value Min within the window; calculate the median Median within the window.

[0031] Judgment condition: If Min < Center < Max and Min < Median < Max, then it is considered that the current pixel has no noise and its original value can be retained; otherwise, it is considered that this pixel is noise and is replaced with the median, where Center is the pixel value being processed currently.

[0032] S5. If expanding the window still cannot effectively remove the noise, the algorithm will continue to increase the window until the maximum window size is reached or the denoising effect meets the expectation. If the noise still cannot be removed after the window is expanded to the maximum limit, the window expansion will stop.

[0033] S6. After the filtering process is completed, output the denoised image. Each pixel of the entire image has undergone non-restrictive median filtering processing, removing the noise and retaining more detailed information.

[0034] Further, the edge extraction step is as follows: The denoised image is input into a pre-trained deep learning model. The model performs edge extraction of the pattern according to the learned mapping relationship between the original image and the edge image, extracts edge features at different levels through multiple blocks of the perceptual fusion sub-network, and further optimizes the resolution and clarity of the edge map through the depth sampling sub-network.

[0035] Further, the steps of vectorization processing and saving the output are as follows: After edge extraction, the model identifies the key features of the pattern, converts these edge maps into vector maps through a vectorization algorithm, and saves them digitally. The vector map can remain clear at different sizes, meet the requirements of the digital platform, and provide an interface for downloading or further processing. SVG format files can be scaled losslessly and are suitable for various digital platforms and printing applications.

[0036] The advantages of the present invention compared with the existing technologies are as follows:

[0037] 1. The present invention overcomes the noise interference in the extraction of complex patterns. By implementing denoising preprocessing (non-restrictive median filtering) before image processing, the noise in the image is effectively removed, especially the noise interference at the edges of complex patterns, and important detail information is retained. This processing process avoids the problem that the pattern features are lost or inaccurate in the traditional method when processing low-quality or noisy images.

[0038] 2. The edge extraction based on deep learning in the present invention ensures the accuracy of the pattern. By using a neural network model with deep perceptual fusion (including a perceptual fusion sub-network and a depth sampling sub-network), the edge features in the image can be accurately extracted. Especially through parallel skip connections, the model can effectively avoid losing detail information in the deep network, ensuring the integrity and accuracy of the image edges; introducing parallel skip connections can effectively retain and transmit the edge information extracted by each layer of the network, thus solving the problem that traditional neural networks are prone to losing edge features in deep processing. This design helps to maintain a high quality of edge information in the deep layer of the network, making the finally generated edge pattern clearer and more accurate; in the processing of the depth sampling sub-network, multiple intermediate edge maps are spliced and fused to finally generate a high-quality pattern edge map. This method not only improves the detail expressiveness of the pattern but also better reflects the true features of the pattern. Through the adaptive upsampling process, the finally generated edge map can accurately restore the target pattern.

[0039] 3. The present invention has efficient edge feature preservation and expression. It extracts the edge features of an image through a deep learning algorithm and converts the edge pattern into a high-precision SVG format for preservation through vectorization processing. This process ensures that the extracted patterns are not only clear but also scalable, capable of maintaining high-quality presentation at different sizes, and suitable for the diverse requirements in modern digital platforms.

[0040] 4. The present invention is applicable to various pattern types. It is not only suitable for the digital extraction of specific patterns such as Qiang embroidery but also has good versatility and can be applied to the digital processing of various different types of intangible cultural heritage patterns. By training the network, it can automatically extract and identify the pattern features of different cultural heritages, providing strong support for the protection and inheritance of cultural heritages.

[0041] 5. The present invention has enhanced detail restoration and pattern reproduction capabilities. It can effectively restore the details in the image and optimize the pattern reproduction ability through a deep learning algorithm, avoiding the loss of edge information caused by limitations in computing power or model structure in traditional algorithms. Especially when dealing with complex or detailed patterns, this method can better maintain the fineness and coherence of the patterns. Description of the Drawings

[0042] Figure 1 It is the overall processing flow chart of the present invention. Detailed Embodiments

[0043] In order to make the content of the present invention easier to be clearly understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention.

[0044] Embodiment 1

[0045] The present invention provides a pattern extraction method based on denoising preprocessing, aiming to accurately extract the pattern information in an image through steps such as denoising, edge extraction, pattern recognition, and vectorization processing, and generate a high-precision scalable vector graphic format for preservation. By denoising preprocessing and deep learning algorithms, the edge features of intangible cultural heritage patterns are extracted and preserved. The overall process is as follows:

[0046] A. Training stage: Model training and optimization

[0047] The goal of the training stage is to train a neural network model based on depth perception fusion through a large amount of sample data so that it can accurately extract edge information from the original image. The steps in this part include data preparation, model training, and optimization.

[0048] A1. Data preparation

[0049] Input image data (including the paired graphs of each original image and its corresponding edge image, where the original image is used to provide detailed information about the pattern, and the edge image is used to provide edge features as the desired output). The dataset is divided into a training set, a test set, and a validation set. The training set is used for training the model, the test set is used to verify the performance of the model, and the validation set is used to evaluate the model's performance on new data and tune the hyperparameters.

[0050] A2. Model Structure

[0051] The core of the model is a neural network based on depth-aware fusion, which mainly consists of two major parts: a perception fusion subnet and a depth sampling subnet.

[0052] Perception Fusion Subnet: It consists of five encoder-like output blocks (k-1 to k-5). Each block is composed of a combination of multiple smaller sub-blocks. The main function of the perception fusion subnet is to extract features from the low layer to the high layer, and connect each block and its sub-blocks together through skip connections to ensure that edge information at different levels is retained. The main purpose of the perception fusion subnet is to extract the edge information of the pattern at multiple levels and avoid the loss of edge features through skip connections.

[0053] Among them, block k–1 has 1 sub-block. The sub-block has 1 convolutional layer with a stride of 2 (s2), a kernel size of 3x3, 32 convolutional kernels, and 1 convolutional layer with 64 convolutional kernels, a kernel size of 3x3, performing a conventional convolution operation without changing the stride; block k-2 has two sub-blocks. One sub-block has two convolutional layers with 128 convolutional kernels, a kernel size of 3x3, and a stride of 2, and the other sub-block is a max pooling layer, a 3x3 max pooling layer with a stride of 2. Block k-3 has three sub-blocks. Two of the sub-blocks each have two convolutional layers with 256 convolutional kernels, a kernel size of 3x3, and a stride of 2, and the other sub-block is a max pooling layer; block k-4 has four sub-blocks. Three of the sub-blocks each have two convolutional layers with 512 convolutional kernels, a kernel size of 3x3, and a stride of 2, and the other sub-block is a max pooling layer; block k-5 has three sub-blocks, and all three sub-blocks each have two convolutional layers with 512 convolutional kernels, a kernel size of 3x3, and a stride of 2.

[0054] The feature maps generated by each block are input into an independent depth sampling subnet. Before merging with the main connection, an average operation of the edge connection is performed to generate an intermediate edge pattern map. Finally, these intermediate edge maps are fused and spliced together to form a learning stack. The last step of the network fuses these features to generate the final pattern map.

[0055] Deep Sampling Sub-network: It consists of two blocks (b-1 and b-2), which process feature maps through a conditional stacking structure. Each block contains a convolutional layer and a transposed convolutional layer. This network is responsible for upsampling the extracted intermediate edge map and finally generating an edge map with the same size as the original image.

[0056] b-1 processes the input using a 1×1 convolutional kernel, followed by the ReLU activation function, and then performs a transposed s×s convolution, where the size s of the convolutional kernel is determined by the scale level of the input feature map. b-2 is only activated when the input feature map needs to be scaled from the separable sub-network. This block processes iteratively until the size of the feature map reaches twice the target size. When the condition is met, the feature map is fed back to Block-1;

[0057] A3. Model Training and Optimization

[0058] Using the training set data in the image data provided by A1, feature learning and local information extraction are performed on the network architecture of A2. The provided original image and its corresponding paired edge image are integrated and optimized, and then the model is tested and verified through the test set and validation set to obtain an optimized neural network model based on depth perception fusion and model parameters.

[0059] B. Application Stage: Pattern Extraction and Preservation

[0060] In the application stage, the non-restrictive median filtering technique is first used to denoise the pattern image to be extracted, and then the trained neural network model based on depth perception fusion is used for pattern extraction, and it is converted into a final vector graph format that can be saved through image processing techniques.

[0061] B1. Image Acquisition and Digitization

[0062] According to different application requirements, images can be obtained by shooting with professional photographic equipment (such as photos of embroidered Qiang embroidery objects) to obtain physical images, or digital images can be obtained by designing patterns on online platforms. The collected image data should have sufficient clarity and high resolution to ensure that the details of the pattern can be accurately extracted in subsequent processing. The collected images should be converted into standard digital image formats (such as PNG, JPEG, or TIFF formats) for subsequent processing.

[0063] B2. Denoising Preprocessing

[0064] First, grayscale the image collected in step B1 by performing a weighted average on the red, green, and blue (RGB) color channels of each pixel to obtain its grayscale value. Since impurities during the shooting process increase the noise of the image (for example, in the actual image of Qiang embroidery, the edges of the image may be discontinuous due to embroidery techniques, resulting in salt-and-pepper-like noise points), use non-restrictive median filtering denoising technology.

[0065] Specific steps of the non-restrictive median filtering algorithm:

[0066] 1. Set an initial window size (initially set to 3x3). This window will slide pixel by pixel in the image and calculate the median of each pixel.

[0067] 2. During the filtering process, the algorithm does not restrict the specific size of the window. It determines whether to expand or shrink the window by checking the pixel value changes within the current window. If the noise within the window is strong, the algorithm will increase the window size (for example, expand to 5x5, 7x7, etc.) to better smooth the noise. Each time the window size is expanded, the maximum expansion can reach a fixed upper limit (set to 10x10).

[0068] 3. For each pixel point, calculate the median of all pixel values within its window. The median refers to the middle value after sorting the pixel values in the window. In some cases, if the pixel value within the window differs significantly from the value of the central pixel, then this pixel is considered noise and is replaced with the median.

[0069] 4. Determine whether this pixel is noise based on the pixel value changes within the current window. If the current pixel has a significant difference from the pixels in its neighborhood, then this pixel is considered noise and is replaced with the median. The judgment criterion is usually through the following steps:

[0070] Calculate the maximum value (Max) and minimum value (Min) within the window.

[0071] Calculate the median (Median) within the window.

[0072] Judgment condition: If Min < Center < Max and Min < Median < Max, then it is considered that the current pixel has no noise and its original value can be retained; otherwise, it is considered that this pixel is noise and is replaced with the median, where Center is the pixel value being processed currently.

[0073] 5. If expanding the window still cannot effectively remove the noise, the algorithm will continue to increase the window until the maximum window size is reached or the denoising effect meets the expectation. Generally speaking, if the noise still cannot be removed after the window is expanded to the maximum limit, the window expansion will stop.

[0074] 6. After the filtering process is completed, the denoised image is output. Each pixel of the entire image has undergone unrestricted median filtering, removing noise and retaining more detailed information.

[0075] B3. Edge extraction

[0076] The denoised image is input into the trained deep learning model. The model extracts the edges of the pattern based on the learned mapping relationship between the original image and the edge image. Different levels of edge features are extracted through multiple blocks of the perceptual fusion sub-network, and the resolution and clarity of the edge map are further optimized through the depth sampling sub-network.

[0077] B4. Vectorization processing and saving the output

[0078] After edge extraction, the model identifies the key features of the pattern. Then, these edge maps are converted into vector graphics through a vectorization algorithm and saved digitally. Vector graphics can remain clear at different sizes, meeting the requirements of digital platforms. An interface for downloading or further processing is provided. SVG format files can be scaled losslessly and are suitable for various digital platforms and printing applications.

[0079] Embodiment 2

[0080] In the process of pattern extraction, the present invention selects unrestricted median filtering as the denoising preprocessing technology. By not restricting the window size but adaptively selecting the sliding window size and calculating the similarity weight according to pixel points, noise is removed while details are retained.

[0081] The present invention adopts a neural network architecture of deep perceptual fusion, including a perceptual fusion sub-network and a depth sampling sub-network. Through operations such as multi-layer convolution and skip connections, this network structure effectively extracts and retains edge features in complex patterns, avoiding problems such as edge breakage and noise interference in traditional methods, and can provide continuous and complete edge information as the basis for subsequent processing.

[0082] The present invention uses a deep learning model to extract the edge features of an image and generates a high-precision scalable vector graphic in SVG format through vectorization processing. This processing ensures that the edge information of the image can remain clear and accurate at different sizes.

[0083] The method of the present invention does not rely on a pre-trained model for initialization, but directly extracts features from the training data paired with the original image and the edge image through a self-learning method. This enables the method to adapt to various different types of image data.

[0084] The present invention can effectively preserve traditional art patterns (such as intangible cultural heritages like Qiang embroidery), and display, modify, and share them through digital platforms, providing a new technical path for the inheritance of cultural heritages.

[0085] The method for digital preservation and sharing of cultural heritage based on automated processes such as image preprocessing, edge extraction, and vectorization proposed by the present invention is particularly suitable for the modern inheritance of intangible cultural heritage.

[0086] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of the present invention, design similar structural manners and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. A pattern extraction method based on denoising preprocessing, characterized in that, It includes the following steps: Step 1: Model training and optimization; First, perform data preparation. Input image data, and divide the dataset into a training set, a test set, and a validation set. The training set is used for model training, the test set is used to verify the performance of the model, and the validation set is used to evaluate the model's performance on new data and tune hyperparameters; The core of the model is based on a neural network for depth-aware fusion. The network structure includes: a perception fusion subnetwork and a depth sampling subnetwork; The model uses the provided original image and its paired edge image to perform integration and optimization. The model is tested and verified through the test set and the validation set to obtain an optimized neural network model for depth-aware fusion and model parameters; Step 2: Pattern extraction and saving; First, use non-restrictive median filtering technology to denoise the pattern image to be extracted. Then, use the trained neural network model for depth-aware fusion to perform pattern extraction, and convert it into a final vector graph format that can be saved through image processing technology. The pattern extraction and saving include image acquisition and digitization, denoising preprocessing, edge extraction, and vectorization processing and saving output.

2. The pattern extraction method based on denoising preprocessing according to claim 1, wherein: In Step 1, the input image data includes each original image and its paired edge image. The original image is used to provide detailed information about the pattern, and the edge image is used to provide edge features as the expected output.

3. The pattern extraction method based on denoising preprocessing according to claim 1, wherein: In Step 1, the perception fusion subnetwork includes five encoder-like output blocks, which are output blocks K-1 to K-5. Each output block is composed of a combination of multiple smaller sub-blocks; Among them, block K-1 has one sub-block. The sub-block has a convolutional layer s2 with a stride of 2, a convolutional kernel size of 3x3, 32 convolutional kernels, and a convolutional layer with 64 convolutional kernels, a convolutional kernel size of 3x3, and performs a conventional convolution operation without changing the stride; Block K-2 has two sub-blocks. One sub-block has two convolutional layers with 128 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2. The other sub-block is a max pooling layer, a 3x3 max pooling layer with a stride of 2; K-3 has three sub-blocks. Two of the sub-blocks each have two convolutional layers with 256 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2. The other sub-block is a max pooling layer; K-4 has four sub-blocks. Three of the sub-blocks each have two convolutional layers with 512 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2. The other sub-block is a max pooling layer; K-5 has three sub-blocks, and all three sub-blocks have two convolutional layers with 512 convolutional kernels, a convolutional kernel size of 3x3, and a stride of 2; The feature maps generated by each output block are input into an independent depth sampling subnetwork. Before merging with the main connection, an average operation of the edge connection is performed to generate an intermediate edge pattern map. Finally, these intermediate edge maps are fused and spliced together to form a learning stack. The last step of the network fuses these features to generate the final pattern map.

4. A pattern extraction method based on denoising preprocessing according to claim 1, characterized in that: The deep sampling sub-network includes two output blocks, namely b-1 and b-2. The output blocks process feature maps through a conditional stacking structure. Each block contains a convolutional layer and a transposed convolution. The deep sampling sub-network is responsible for upsampling the extracted intermediate edge map and finally generating an edge map with the same size as the original image; b-1 processes the input using a 1×1 convolutional kernel, followed by the ReLU activation function, and then performs a transposed s×s convolution, where the size s of the convolutional kernel is determined by the scale level of the input feature map; b-2 is only activated when the input feature map needs to be scaled from the separable sub-network. This block will be iteratively processed until the size of the feature map reaches twice the target size. When the condition is met, the feature map is fed back to Block-1.

5. A pattern extraction method based on denoising preprocessing according to claim 1, characterized in that: In step two, the steps of image acquisition and digitization include: According to different application requirements, the image is obtained by taking a physical image through a professional photography device or by designing a pattern through an online platform to obtain a digital image; The acquired image data has high definition and high resolution to ensure that the subsequent processing can accurately extract the details of the pattern; The acquired image is converted into a standard digital image format for subsequent processing.

6. A pattern extraction method based on denoising preprocessing according to claim 1, characterized in that: The steps of the noise removal preprocessing include: grayscale processing the acquired image, calculating the weighted average of the red, green, and blue color channels of each pixel to obtain its grayscale value. The impurities in the image during the shooting process increase the noise of the image, and a non-restrictive median filtering denoising technique is used.

7. A pattern extraction method based on denoising preprocessing according to claim 6, characterized in that: The non-restrictive median filtering algorithm includes the following steps: S1. Set an initial window size, initially set to 3x3. The window will slide pixel by pixel in the image and calculate the median of each pixel; S2. During the filtering process, the algorithm does not limit the specific size of the window. It decides whether to expand or shrink the window by checking the change of pixel values within the current window; If the noise within the window is strong, the algorithm will increase the size of the window to better smooth the noise. Each time the window size is expanded, it can be expanded up to a fixed upper limit, set to 10x10; S3. For each pixel point, calculate the median of all pixel values in its window. The median refers to the middle value after sorting the pixel values in the window. In some cases, if the pixel value within the window differs significantly from the value of the central pixel, then this pixel is considered noise and is replaced with the median; S4. Determine whether this pixel is noise based on the change of pixel values within the current window. If the current pixel has a significant difference from the pixels in its neighborhood, then this pixel is considered noise and is replaced with the median. The judgment criterion is through the following steps: Calculate the maximum value Max and the minimum value Min within the window; Calculate the median Median within the window; Judgment condition: If Min < Center < Max and Min < Median < Max, then it is considered that the current pixel has no noise and its original value can be retained; Otherwise, it is considered that this pixel is noise and is replaced with the median, where Center is the pixel value being processed currently; S5. If the extended window still cannot effectively remove noise, the algorithm will continue to increase the window until the maximum window size is reached or the denoising effect meets the expectation; if the noise still cannot be removed after the window is extended to the maximum limit, the window extension will be stopped. S6. After the filtering process, the denoised image is output. Each pixel of the entire image has undergone unrestricted median filtering to remove noise and retain more detailed information.

8. A pattern extraction method based on denoising preprocessing according to claim 1, characterized in that: The edge extraction step is as follows: The denoised image is input into the trained deep learning model. The model performs edge extraction of the pattern based on the learned mapping relationship between the original image and the edge image, extracts edge features at different levels through multiple blocks of the perception fusion sub-network, and further optimizes the resolution and clarity of the edge map through the depth sampling sub-network.

9. A pattern extraction method based on denoising preprocessing according to claim 1, characterized in that: The steps of vectorization processing and saving the output are as follows: After edge extraction, the model will identify the key features of the pattern, convert these edge maps into vector graphics through the vectorization algorithm, and save them digitally. The vector graphics can remain clear at different sizes, meet the requirements of the digital platform, provide an interface for downloading or further processing, and SVG format files can be scaled losslessly and are suitable for various digital platforms and printing applications.

Images