Image line draft generation method and system based on dynamic parameter sliding window and dual-mode arbitration

Abstract

The invention discloses an image line draft generation method and system based on a dynamic sliding window and dual-mode arbitration, and mainly solves the problems that in the prior art, details are lost, the processing time of a high-resolution image is too long, and the classification accuracy of detail areas is low. According to the implementation scheme, the method comprises the following steps: calculating a window size and a sliding step length according to indexes of resolution, local texture complexity and color contrast of an input image, and carrying out sliding window partitioning on the input image; performing semantic segmentation on each segmented sub-window to obtain a segmentation result and a pixel-level confidence coefficient of each sub-window image; carrying out weighted fusion on pixels of the sliding window overlapping region to generate an initial line draft segmentation result, and carrying out parallel optimization of image segmentation and clustering on all image detail regions of the initial line draft segmentation result; and re-optimizing the unoptimized detail area through artificial feedback arbitration to generate a final line draft segmentation result. According to the method, detail loss caused by image compression can be avoided, the high-resolution image processing time is shortened, the classification accuracy of complex detail areas is improved, and the method can be used for animation production, artistic design and medical image analysis.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a method and system for generating image line drawings. It can be used in animation production, art design, and medical image analysis. Background Technology

[0002] Image line art generation is a key technology in computer vision and image processing, widely used in animation production, art design, and medical image analysis. Its core objective is to extract contour features and structural information from images using algorithms to generate line art with high precision and visual appeal. Traditional methods are mainly divided into two categories: methods based on manually designed features and methods based on deep learning. The former relies on edge detection algorithms such as Canny and Sobel, as well as image segmentation techniques such as thresholding and region growing. While it can extract basic contour information, it often suffers from edge breakage and loss of detail when processing complex scenes or high-resolution images. The latter uses convolutional neural networks (CNNs) to achieve end-to-end feature learning, improving line art generation accuracy through large-scale data training. However, it is limited by the memory usage of high-resolution images (a typical end-to-end CNN model requires more than 12GB of memory to process a single 4K image) and computational efficiency, resulting in bottlenecks such as slow processing speed and blurred edges.

[0003] Patent application number 202410422107.9 discloses an unsupervised neighborhood classification superpixel generation method and system incorporating guided filtering. It achieves pixel neighborhood classification through a lightweight model, using a U-Net network for mapping and downsampling for image compression to improve computational efficiency. However, this method suffers from irreversible loss of high-frequency detail information due to image compression during superpixel generation. Furthermore, the lack of sufficient consideration of spatial continuity constraints in the neighborhood classification mechanism leads to fragmentation of segmentation results in areas with complex textures, affecting classification accuracy.

[0004] Patent application number 202410272406.9 discloses a multi-source remote sensing image classification method based on multi-level feature fusion. It employs the Simple Linear Iterative Clustering (SLIC) algorithm for pixel segmentation and fuses local texture features with global semantic features through feature concatenation. However, this method is limited by the SLIC algorithm's O(n) time complexity. 2 The time complexity of this approach increases significantly, leading to a substantial increase in processing time. More importantly, its uniform stitching-based feature extraction and fusion strategy struggles to adapt to the differences in land cover types within remote sensing images, resulting in low classification accuracy for complex and detailed regions. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing an image line drawing generation method and system based on dynamic parameter sliding window and dual-mode arbitration, so as to avoid the loss of details caused by image compression, reduce the processing time of high-resolution images, and improve the classification accuracy of complex detail areas.

[0006] The core technology for achieving the objective of this invention is to construct an architecture for a dynamic sliding window segmentation strategy and a dual-mode collaborative decision-making mechanism, the implementation of which includes:

[0007] 1. A method for generating image line drawings based on dynamic parameter sliding window and dual-mode arbitration, characterized in that it includes:

[0008] (1) Based on the resolution (W) of the input image 原 H 原 The image is divided into sliding window blocks by calculating the local texture complexity and color contrast of the image as two-dimensional indicators, the window size W and the sliding step S.

[0009] (2) Perform open-world semantic segmentation on each segmented sub-window and record the pixel-level confidence score of each sub-window. i (x,y);

[0010] (3) Pixel-level confidence score based on its sub-window i (x,y) performs weighted fusion on the pixels in the overlapping area of ​​the sliding window to generate the initial line drawing segmentation result;

[0011] (4) All image detail regions in the initial line drawing are optimized in parallel using graph cut algorithm and clustering algorithm;

[0012] (5) Calculate the segmentation overlap rate η between the graph cut optimization result and the clustering optimization result of the same detail region, and judge it accordingly:

[0013] If η > 75%, then update the detail region in the initial line drawing with either the graph cut optimization or the clustering optimization result.

[0014] Conversely, manual feedback is activated to receive users' arbitration decisions on disputed areas, and the line drawing details are updated based on the user's selection.

[0015] After all detailed areas have been optimized, the final line art segmentation result is generated.

[0016] 2. An image line drawing generation system based on dynamic parameter sliding window and dual-mode arbitration, characterized in that it includes:

[0017] A dynamic sliding window segmentation module is used to segment the input image into blocks according to the method described in claims 1 and 2;

[0018] The sub-window segmentation module is used to call an open-world semantic segmentation model to process each sub-window in parallel, and output the segmentation results and pixel confidence scores. i (x,y);

[0019] The sub-line drawing confidence fusion module is used to score overlapping region pixels. i (x,y) weighted fusion is used to output the initial line drawing;

[0020] The dual-mode optimization module is used to perform dual-mode optimization on image details.

[0021] The human feedback arbitration module is used to receive point-selection or stroke-selection arbitration commands input by users through the brush tool, optimize the area selected by the user, and output the final line drawing.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] Firstly, this invention employs a dynamic sliding window block strategy based on image resolution, local texture complexity, and color contrast, combined with pixel-level confidence weighted fusion of overlapping regions, which avoids the loss of image details due to global image compression.

[0024] Secondly, by combining texture complexity and color contrast to dynamically adjust the window size and sliding step size, this invention achieves adaptive and efficient processing of high-resolution images, thereby reducing the processing time for high-resolution images.

[0025] Third, this invention improves the accuracy of fine-grained region segmentation in complex scenes by performing dual-mode parallel optimization of graph cut and clustering in image detail regions and introducing a manual feedback arbitration mechanism based on overlap rate judgment. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the image line drawing generation method based on dynamic sliding window and dual-mode arbitration of the present invention.

[0027] Figure 2 This is a schematic diagram of the dynamic sliding window block processing in the method of the present invention;

[0028] Figure 3 This is a flowchart of the manual feedback arbitration mechanism in the method of this invention;

[0029] Figure 4 This is a block diagram of the image line drawing generation system based on dynamic sliding window and dual-mode arbitration of the present invention;

[0030] Figure 5 This is a simulation diagram of the line drawing generation of the input image according to the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should all fall within the protection scope of the present invention.

[0032] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

[0033] Example 1: Image Line Drawing Generation Method Based on Dynamic Sliding Window and Dual-Mode Arbitration

[0034] Reference Figure 1 The implementation steps of this example include the following:

[0035] Step 1: Calculate the window size W and sliding step S based on the resolution (W0, H0) of the input image, the local texture complexity of the image, and the color contrast.

[0036] 1.1) Using a×a pixels as the basic window unit, slide through the entire input image and calculate the variance σ of the pixels within each a×a window. 2 and information entropy E grad :

[0037]

[0038] in, The base window size, where W0 is the width of the input image and H0 is the height of the input image. x represents the average grayscale value of the pixels within the window. i Let m be the grayscale value of the i-th pixel within the window. i Let μ be the gradient magnitude of the i-th pixel within the window. m This is the average gradient magnitude of all pixels within the window.

[0039] 1.2) Based on the variance σ 2 and information entropy E grad Calculate the window size W and sliding step size S for different regions:

[0040]

[0041] Where ρ is the window overlap ratio adjustment factor.

[0042] 1.3) Divide the input image into blocks according to the window size W and the sliding step S:

[0043] Starting from the top left corner of the input image, the sub-window is cropped using the W calculated within the a×a window at the top left corner as the side length of the rectangle. After cropping, the S is used as the sliding distance of the a×a window, and the a×a window is shifted to the right. The values ​​of W and S calculated for the current a×a window are updated.

[0044] Crop and pan again, wrapping to the rightmost edge, until the entire image is divided into multiple sub-windows, as shown below. Figure 2 As shown, by Figure 2 It can be seen that as the variance σ of pixels within the a×a window in the highly complex region increases... 2 and information entropy E grad As the window size W increases, the sliding step size S decreases, resulting in denser blocks, as shown in the attached diagram. Figure 2 Sub-windows numbered 1-9; conversely, the blocks are more sparse, as shown in the attached figure. Figure 2 Sub-windows 45-49;

[0045] 1.4) The sub-windows of the block are grouped according to the window size W. Windows of the same size are grouped together, and the sub-windows of the same group are stacked according to the batch dimension in the computer image to facilitate subsequent parallel computing.

[0046] Step 2: Parallel computation is performed on the sub-window group using a general semantic segmentation model to obtain the content segmentation results and pixel-level confidence scores of each sub-window image. i (x,y).

[0047] The general semantic segmentation model includes a fully convolutional neural network architecture, such as UNet, DeepLabV3+ and its variants, and a convolutional-Transformer hybrid architecture, such as Swin-UNet, Trans-UNet and its variants. It features a batch tensor input interface and a dual-branch output layer. The batch tensor input interface receives a four-dimensional tensor of shape [batch, 3, W, W]. In the dual-branch output layer, branch 1 outputs a binary segmentation mask of shape [batch, 1, W, W], and branch 2 outputs pixel confidence scores of shape [batch, 1, W, W]. i (x,y).

[0048] This step utilizes the model's batch processing tensor input interface to receive sub-window group tensors for parallel computation, and leverages the model's dual-branch output layer to obtain the segmentation results and pixel confidence scores for the sub-windows. i (x,y).

[0049] Step 3, based on pixel-level confidence score i (x,y) yields the initial line art.

[0050] 3.1) Let Scorex (x,y) represents the confidence score of the segmentation result of the x-th sub-window at pixel coordinates (x,y). other Mask is used to assess the confidence level of segmentation results for other child windows at pixel coordinates (x, y). x (x,y) represents the segmentation result of the x-th sub-window at pixel coordinates (x,y);

[0051] 3.2) Determine Score x (x,y)>Score other Determine if (x,y)+0.2 is true, and what is the segmentation result at pixel coordinates (x,y)?

[0052] If true, then use Mask directly. x (x,y) is the segmentation result Fin(x,y) at pixel coordinates (x,y);

[0053] If this condition is not met, then the segmentation results of all sub-windows covering this coordinate are weighted and fused to obtain the segmentation result Fin(x,y) at pixel coordinates (x,y):

[0054]

[0055] BinaryMask() is a binarization function. When the result of the fused pixels in the overlapping area of ​​the sliding window is ≥0.5, the output is 1; otherwise, the output is 0.

[0056] 3.3) The overlapping and non-overlapping regions in the fusion segmentation result Fin(x,y) are stitched together in the order of image block division to obtain the initial line drawing.

[0057] Step 4: Optimize the initial line art.

[0058] Existing methods for optimizing initial line art include graph cut algorithms, clustering algorithms, edge detection algorithms, and region growing algorithms. This example uses, but is not limited to, graph cut algorithms and clustering algorithms.

[0059] The graph cut algorithm is the Felzenszwalb algorithm in OpenCV, which can automatically segment images based on the graphical structure in the image;

[0060] The clustering algorithm is the K-means++ algorithm in OpenCV, which can automatically segment images based on color differences in the image.

[0061] This step includes:

[0062] 4.1) Obtain the edge response map of the input image using the Canny edge detection algorithm, and calculate the edge response intensity M based on the edge response map of the input image:

[0063]

[0064] Among them, G x G represents the gradient of the image along the horizontal axis. y This represents the gradient of the image along the vertical axis.

[0065] 4.2) The connected regions formed by pixels with edge response intensity M > 200 are considered as image detail regions;

[0066] 4.3) Obtain the graph cut segmentation mask of all detail regions of the input image through the graph cut algorithm, and obtain the cluster segmentation mask of all detail regions of the input image through the clustering algorithm;

[0067] 4.4) Use the findContours() algorithm in OpenCV to extract the boundaries of the above graph cut segmentation mask and cluster segmentation mask, and use them as the graph cut optimization line drawing results and cluster optimization line drawing results of the detail region, respectively;

[0068] 4.5) For each detailed region, calculate the overlap rate η between the graph cut optimization line drawing result and the clustering optimization line drawing result:

[0069]

[0070] Where A represents the set of line drawing pixels in the graph cut optimization line drawing result, B represents the set of line drawing pixels in the clustering optimization line drawing result, |A∩B| represents the number of overlapping pixels at the boundaries of the two optimized line drawings, and |A∪B| represents the total number of non-overlapping pixels at the upper boundaries of the two optimized line drawings;

[0071] 4.6) Optimize the current line art result based on the overlap rate η:

[0072] If η > 75%, then choose to update the corresponding region in the initial line drawing by optimizing the line drawing result with graph cut or clustering.

[0073] If η≤75%, then highlight the detailed area on the input image and display it on the human-computer interaction interface, and proceed to step 5.

[0074] Step 5: Perform manual feedback arbitration for detailed areas with an overlap rate η≤75%.

[0075] Reference Figure 3 The implementation of this step includes the following:

[0076] 5.1) Users select a detail area with an overlap rate η≤75% on the labeled human interface by using dots or strokes, and display the cut optimization line drawing and cluster optimization line drawing of the detail area side by side on the human-computer interaction interface.

[0077] 5.2) Based on the graph cut optimization line drawing and cluster optimization line drawing displayed on the human-computer interaction interface, the user selects the one with the better subjective effect to update the initial line drawing;

[0078] 5.3) Repeat steps 5.1) and 5.2) until all detail areas with an overlap rate η≤75% have been manually arbitrated, and then output the final line drawing.

[0079] Example 2: Image Line Drawing Generation System Based on Dynamic Sliding Window and Dual-Mode Arbitration

[0080] Reference Figure 4 This example includes: a dynamic sliding window segmentation module 1, a sub-window segmentation module 2, a sub-line drawing confidence fusion module 3, a dual-mode optimization module 4, and a human feedback arbitration module 5. The dual-mode optimization module 4 includes: a detail region segmentation sub-module 41, a graph cut optimization sub-module 42, a clustering optimization sub-module 43, a result overlap calculation sub-module 44, and a result optimization sub-module 45. The human feedback arbitration module 5 includes: a human-computer interaction sub-module 51, a result reading sub-module 52, and an arbitration optimization sub-module 53. The working principle of the entire system is as follows:

[0081] The dynamic sliding window segmentation module 1 is used to segment the input image into blocks and output the segmented sub-windows to the sub-window segmentation module 2.

[0082] The sub-window segmentation module 2 is used to call the open-world semantic segmentation model to process the segmented sub-windows in parallel, and obtain the sub-window segmentation results and pixel confidence scores. i (x,y) is then transmitted to the sub-line drawing confidence fusion module 3;

[0083] The sub-line drawing confidence fusion module 3 is used to score the pixels in the overlapping areas of the sub-window segmentation results according to their scores. i (x,y) weighted fusion is used to obtain the initial line drawing, and the initial line drawing is output to the dual-mode optimization module 4 and the human feedback arbitration module 5 respectively;

[0084] The dual-mode optimization module 4 is used to perform dual-mode optimization on the detail parts of the input image. The detail region segmentation submodule 41 segments the image detail regions from the input image and outputs these regions to the graph cut optimization submodule 42 and the clustering optimization submodule 43. The graph cut optimization submodule 42 performs secondary segmentation on the input image detail regions based on a graph cut algorithm to obtain graph cut boundary lines and outputs these lines to the result overlap calculation submodule 44. The clustering optimization submodule 43 performs secondary segmentation on the image detail regions based on a clustering algorithm. Obtain the cluster boundary line drawing and output it to the result overlap calculation submodule 44; the result overlap calculation submodule 44 calculates the overlap rate η between the graph cut boundary line drawing and the cluster boundary line drawing, and outputs the overlap rate η, the graph cut boundary line drawing, and the cluster boundary line drawing to the result optimization submodule 45; the result optimization submodule 45 optimizes the current line drawing result according to the overlap rate η, updates the initial line drawing with graph cut boundary line drawing or cluster boundary line drawing for η>75%, and transmits the graph cut boundary line drawing and cluster boundary line drawing for η≤75% to the manual feedback arbitration module 5;

[0085] The manual feedback arbitration module 5 is used to optimize the user-selected area and output the final line drawing. The human-computer interaction submodule 51 provides a universal arbitration command input interface on both PC and mobile terminals, obtains the user's arbitration command, and outputs it to the result reading submodule 52 and the arbitration optimization submodule 53. The result reading submodule 52 selects a detailed area according to the arbitration command and reads the graph cut boundary line drawing and cluster boundary line drawing of that area and outputs them to the arbitration optimization submodule 53. The arbitration optimization submodule 53 selects the graph cut boundary line drawing and cluster boundary line drawing according to the arbitration command, updates the current line drawing with the line drawing selected by the user, and outputs the final line drawing after iterative processing.

[0086] It should be noted that the above functional modules can be implemented, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as program instruction products. A program instruction product includes one or a set of program instructions. When the program instructions are loaded and executed on a computer, the described process or function is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Program instructions can be stored in a computer-readable and writable storage medium, or transferred from one computer's readable and writable storage medium to another. Program instructions can be transferred between computers, mobile phones, and other network-connected devices via a network.

[0087] The direct coupling or communication connections between the modules shown or discussed in this embodiment can be achieved through indirect coupling or communication connections via interfaces, devices, or modules. The various functional modules and sub-modules in this embodiment can dynamically reside within a single processing unit, or each module can exist physically independently, or two or more modules can dynamically reside within a single processing unit. When these dynamic components are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. This storage medium can be a memory, disk, or optical disc, etc.

[0088] The effectiveness of this invention can be further illustrated by the following simulation results:

[0089] The method of this invention is deployed as a network service on a server equipped with an NVIDIA 3090 graphics card. A public network interface is set up on the server, and a computer with internet access connects to the network service using browser software. Upon entering the human-computer interaction interface, a 4K image containing complex content is input into the system through the input image window. The system automatically outputs an initial line drawing in the segmented image. Then, clicking the detail processing button allows for manual arbitration. After arbitration, the final line drawing is output in the line drawing image window. The result is as follows. Figure 5 As shown, where: Figure 5 (a) is the input image. Figure 5 (b) is the initial line art image. Figure 5 (c) shows the optimization result of the graph cut algorithm. Figure 5 (d) shows the optimization result of the clustering algorithm. Figure 5 (e) Final line art image.

[0090] from Figure 5 As can be seen, the present invention can optimize the detailed regions of the initial line drawing image generated from the input image through graph cut algorithm and clustering algorithm, and finally output a line drawing image with rich details and clear outlines through human feedback arbitration, effectively preserving the complex structural information of the original image.

Claims

1. A method for generating image line drawings based on dynamic parameter sliding window and dual-mode arbitration, characterized in that, include: (1) Based on the resolution (W0,H0) of the input image, the local texture complexity and color contrast of the image, calculate the window size W and the sliding step S, and divide the input image into sliding window blocks. (2) Perform open-world semantic segmentation on each segmented sub-window and obtain the segmentation results and pixel-level confidence scores of each sub-window image. i (x,y); (3) Pixel-level confidence score based on its sub-window i (x,y) performs weighted fusion on the pixels in the overlapping area of ​​the sliding window to generate the initial line drawing segmentation result; (4) All image detail regions in the initial line drawing are optimized in parallel using graph cut algorithm and clustering algorithm; (5) Optimize the details that failed to be optimized again through manual feedback arbitration. After all the details are optimized, the final line drawing segmentation result is generated.

2. The method according to claim 1, characterized in that, The calculation of window size W and sliding step size S in (1) to divide the input image into sliding window blocks includes the following: 1a) Calculate the window size W and sliding step S based on the resolution and local texture complexity index of the input image: Where, σ 2 Let the variance of the a×a pixel window in the input image be . W0 is the width of the input image, H0 is the height of the input image, and E... grad ρ is the gradient direction entropy of the input image; ρ is the window overlap ratio adjustment factor. 1b) Divide the input image into sliding window blocks according to the calculated window size W and sliding step size S.

3. The method according to claim 1, characterized in that, The implementation of open-world semantic segmentation for each sub-window after segmentation in (2) includes: When generating sub-windows from blocks, group sub-windows of different sizes together and stack sub-windows of the same size. A general semantic segmentation model is used to perform parallel computation on the sub-window group to obtain the content segmentation results of each sub-window image and its pixel-level confidence.

4. The method according to claim 1, characterized in that, In (3), the pixel-level confidence score based on its sub-window is used. i The weighted fusion of pixels in the overlapping region of the sliding window (x, y) is calculated using the following formula: Where Fin(x,y) represents the final segmentation result at pixel (x,y) in the overlapping region. BinaryMask() is a binarization function. It outputs 1 when the merged pixel value in the overlapping area of ​​the sliding window is ≥0.5, and 0 otherwise. Mask x (x,y) represents the segmentation result of the x-th sub-window at pixel coordinates (x,y). Score x (x,y) represents the confidence level of the segmentation result of the x-th sub-window at pixel coordinates (x,y). Score other This indicates the confidence level of the segmentation results of other child windows at pixel coordinates (x, y).

5. The method according to claim 1, characterized in that, In step (4), all image detail regions in the initial line drawing are optimized in parallel using graph cut algorithm and clustering algorithm. The implementation includes: 4a) Calculate the edge response map of the initial line drawing using an edge detection algorithm, and identify the connected regions formed by pixels with edge response intensity greater than a preset threshold as the image detail regions; 4b) Use the Felzenszwalb graph cut algorithm and K-means++ clustering algorithm for secondary segmentation of the detail region. 4c) Calculate the segmentation overlap rate η between the graph cut optimization result and the clustering optimization result in the same detail region. The segmentation overlap rate η is the ratio of the number of intersection pixels to the number of union pixels of the graph cut optimization result region and the clustering optimization result region in that detail region, as shown in the following formula: Where A represents the region of the graph cut optimization result, B represents the region of the clustering optimization result, |A∩B| represents the number of intersection pixels, and |A∪B| represents the number of union pixels.

6. The method according to claim 1, characterized in that, The activation of the manual feedback system in (5), which receives users' arbitration decisions on the disputed area, includes the following implementation: 6a) Supports a closed-loop optimization process of "user selection → algorithm optimization → result feedback → parameter correction". 6b) According to the method of claim 1, the user selects a detail area with η>75% using a brush, and two optimization results are presented in real time, with the user deciding the final result.

7. An image line drawing generation system based on dynamic parameter sliding window and dual-mode arbitration, characterized in that, include: The dynamic sliding window segmentation module is used to segment the input image into blocks; The sub-window segmentation module is used to call an open-world semantic segmentation model to process the segmented sub-windows in parallel, and output the sub-window segmentation results and pixel confidence scores. i (x,y); The sub-line drawing confidence fusion module is used to score the pixels in the overlapping areas of the sub-window segmentation results. i (x,y) weighted fusion, output the initial line drawing; The dual-mode optimization module is used to perform dual-mode optimization on the details of the input image. The human feedback arbitration module is used to receive point-selection or stroke-selection arbitration commands input by users through the brush tool, optimize the area selected by the user, and output the final line drawing.

8. The system according to claim 7, characterized in that, The dual-mode optimization submodule includes: The detail region segmentation submodule is used to segment the image detail region from the input image; The graph cut optimization submodule is used to perform secondary segmentation of the image detail region based on the graph cut algorithm and obtain the boundary line drawing; The clustering optimization submodule is used to perform secondary segmentation of the image detail region based on the clustering algorithm and obtain the boundary line drawing; The result overlap calculation submodule is used to calculate the overlap rate η between the boundary line drawing of the graph cut optimization submodule and the boundary line drawing of the clustering optimization submodule in the same image detail region. The result optimization submodule is used to optimize the current line drawing result based on the overlap rate η. It can select either the graph cut optimization submodule or the clustering optimization submodule to update the corresponding area in the current line drawing, or optimize the area selected by the user through the manual feedback arbitration submodule to output the final line drawing.

9. The system according to claim 7, characterized in that, The manual feedback arbitration submodule includes: The human-computer interaction submodule is used to provide a universal arbitration instruction input interface on both PC and mobile devices; The result reading submodule is used to read the two optimization results generated by the dual-mode optimization submodule for the user-selected area; The arbitration optimization submodule is used to select between two optimization results read by the result reading submodule based on the instructions input by the user through the human-computer interaction submodule, and replace the corresponding detail area in the current line drawing result with the result selected by the user.

Images