A high-resolution image segmentation and transparency annotation method and system

Abstract

The application relates to the technical field of machine vision, and discloses a high-resolution image segmentation and transparency labeling method and system, which comprises the following steps: S1: inputting a high-resolution original image to be labeled, calling at least two pre-stored models for reasoning to generate a corresponding number of transparency maps; S2: visualizing all the transparency maps and judging whether a transparency map meeting preset accuracy requirements exists, if yes, directly saving the transparency map as a GT label; if no, entering step S3; S3: optimizing the error types in the transparency map by combining interactive processing corresponding to the error types and a multi-round iteration strategy; and S4: uniformly fusing the optimized results to generate a final GT label and save the final GT label to a training set, so that the method is suitable for high-precision data set making in various scenes such as person segmentation, semi-transparent object segmentation and hair segmentation, is convenient to operate, has strong adaptability, and can well meet the demand of the computer vision field for high-quality labeled data.

Description

Technical Field

[0001] This application relates to the technical field of machine vision, and in particular to a high-resolution image segmentation and transparency annotation method and system. Background Technology

[0002] In the field of computer vision, image segmentation and matting techniques are increasingly widely used. Their core relies on a large amount of training data with precise alpha matte labels to support subsequent model training and optimization. However, existing techniques face several unresolved issues in the segmentation and alpha matting of high-resolution images (e.g., 4K and above, resolution ≥3840×2160): First, manual annotation is extremely inefficient and costly. Transparency labels require pixel-level precision, and for detailed areas such as hair strands, fur, semi-transparent fabrics, and glass, the workload of manual pixel-by-pixel correction is enormous. Moreover, the annotation results are easily affected by subjective factors, making it difficult to guarantee consistency. For high-resolution images such as 4K, the number of pixels usually reaches tens of millions, and manual annotation takes a long time, which cannot meet the needs of large-scale dataset production. In addition, existing methods mostly rely on single image segmentation or matting models, and the generalization ability of a single model is limited. In complex background scenes, problems such as foreground misjudgment, over-identification, or missed detection are prone to occur, resulting in insufficient reliability of the annotation results. Summary of the Invention

[0003] In order to provide a labeling method and system that can combine the advantages of multiple models, adapt to high-resolution images, and have targeted local optimization and iterative correction capabilities, this application provides a high-resolution image segmentation and transparency labeling method and system.

[0004] Firstly, this application provides a high-resolution image segmentation and transparency annotation method, employing the following technical solution: A high-resolution image segmentation and transparency annotation method includes the following steps: S1: Input the high-resolution original image to be labeled, call at least two pre-stored models for inference, and generate the corresponding number of transparency maps; S2: Visualize all transparency images and determine if there is a transparency image that meets the preset accuracy requirements. If so, save it directly as a GT tag; otherwise, proceed to step S3. S3: For the error types in the transparency map, optimize by combining the corresponding interactive processing and multi-round iteration strategy; S4: Unify and merge the optimized results to generate the final GT labels and save them to the training set.

[0005] Optionally, in step S3, the error types include subject recognition error, foreground omission, multiple recognition, and insufficient accuracy in detail areas; The interaction processing and multi-round iteration strategies corresponding to the aforementioned error types include: Corresponding to subject recognition errors and foreground omissions: interactively select the target subject area, re-perform multi-model inference for that area, select the optimal result, and then fuse it with the uncropped part of the whole image through a weighted transparency fusion algorithm, and use an edge smoothing algorithm to process the stitching edges; Corresponding to multi-recognition: The interaction will paint the misjudged area with the preset background color, and perform multi-model inference again. The model will exclude the background interference area based on the color difference. To address the insufficient precision in detailed areas: interactively crop the detailed areas, perform high-resolution inference on the cropped areas to generate a fine transparency map, and then fuse it with the preliminary results of the full image.

[0006] Optionally, the method of fusing the weighted transparency fusion algorithm with the uncropped portion of the full image includes: Fusion formula: ; Where j is the index of the pre-stored model being called. The transparency value of the i-th pixel after merging. Let j be the weight of the j-th model at the i-th pixel. The transparency value of the i-th pixel output by the j-th model; During fusion, the model weights corresponding to the local optimum are used first, and the original weight allocation is maintained for the uncropped parts of the whole image.

[0007] Optionally, the weight The calculation method is as follows: ,in is the confidence score of the j-th model at the i-th pixel, output during the model inference process, representing the reliability of the model's transparency prediction result for that pixel.

[0008] Optionally, the edge smoothing algorithm employs a bilateral filtering algorithm to smooth the fused edge region.

[0009] Optionally, the multi-round iteration strategy includes: The iteration is initiated based on the iteration initiation conditions. Based on the optimization results of the previous round and combined with the low-quality region identification algorithm, a new heatmap of low-quality regions is generated. At the same time, the error regions that were not marked as resolved in the previous round are associated with the heatmap and a list of regions to be processed in this round is compiled. The visual interface displays a list of areas to be processed in this round and a heat map, ultimately confirming the scope of processing in this round. Based on the error type of the area to be processed in this round, the corresponding interactive processing and multi-round iteration strategy are invoked, and the execution process is consistent with the single-round optimization. After optimization, the accuracy index of this round of optimization results is calculated, and a verification report is generated, showing the number of regions processed in this round, the accuracy comparison before and after optimization, and the status of remaining unresolved regions, so that users can determine whether further iteration is needed.

[0010] Optionally, the low-quality region identification algorithm includes: ; in, For the quality estimate of the i-th pixel, Let be the average transparency of n models at the i-th pixel; n is the total number of models involved in the inference. when At that time, the pixel belongs to a low-quality area. This is a preset threshold.

[0011] Optionally, the pre-stored model includes BiRefNet, InSpyreNet, BEN, and RMBG models; The original high-resolution image has a minimum resolution of 4K. During inference, the original resolution of the image is maintained or the original resolution of local areas is maintained, without overall downsampling and compression.

[0012] Secondly, this application provides a high-resolution image segmentation and transparency annotation system, which adopts the following technical solution: A high-resolution image segmentation and transparency annotation system, comprising: Data input module: used to receive the high-resolution raw image to be labeled and transmit it to the multi-model inference module; Multi-model inference module: Pre-stores at least two pre-stored models, used to call at least two pre-stored models for inference, and generate a corresponding number of transparency maps; Visualization filtering module: This module is used to visualize all the transparency maps output by the multi-model inference module and provides an accuracy judgment function to filter preliminary results that meet the requirements. Interactive operation module: Provides interactive tools, including selection and cropping tools, smudge tools, and area selection tools, and provides functions for specifying processing areas and marking error types; Iterative optimization module: Used to execute corresponding optimization strategies based on the specified error type and processing area, and supports multiple rounds of iterative correction; Label output module: Used to save the transparency map finally generated by the iterative optimization module as GT labels and store it in the training set database.

[0013] In summary, this application includes at least one of the following beneficial technical effects: This application fully integrates the advantages of various models through multi-model inference, avoiding the limitations of a single model in high-resolution image processing and effectively improving the overall reliability of the annotation results. Dedicated interactive optimization strategies designed for different error types accurately address issues such as subject recognition errors, multiple recognitions, and insufficient detail precision. Combining high-resolution local inference with multi-resolution fusion, it significantly improves the annotation accuracy of detailed regions while ensuring overall image annotation efficiency. The automatic low-quality region recognition function reduces the workload of manual review, while the multi-stage iterative optimization mechanism gradually corrects remaining problems, significantly reducing the time cost and difficulty of manual refinement. The generated high-precision ground truth (GT) labels can be directly used for model training, effectively improving the performance of subsequent models. It is suitable for creating high-precision datasets in various scenarios such as human segmentation, semi-transparent object segmentation, and hair segmentation. It is easy to operate, highly adaptable, and can well meet the needs of the computer vision field for high-quality labeled data. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the high-resolution image segmentation and transparency annotation method in this application. Detailed Implementation

[0015] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0016] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0017] This application discloses a high-resolution image segmentation and transparency annotation method, which aims to solve the problems of insufficient accuracy of single models, high cost of manual annotation, and lack of targeted optimization and iteration capabilities of interactive tools in the existing high-resolution image segmentation and transparency annotation process. It provides a method and system that combines multi-model inference, interactive local optimization, high-resolution specialized processing and multi-stage iterative correction to efficiently generate high-precision transparency labels and meet the needs of large-scale, high-quality dataset production.

[0018] This application discloses a high-resolution image segmentation and transparency annotation system, including: Data input module: used to receive the high-resolution raw image to be labeled and transmit it to the multi-model inference module; Multi-model inference module: Pre-stores at least two pre-stored models, used to call at least two pre-stored models for inference, and generate a corresponding number of transparency maps; Visualization filtering module: This module is used to visualize all the transparency maps output by the multi-model inference module and provides an accuracy judgment function to filter preliminary results that meet the requirements. Interactive operation module: Provides interactive tools, including selection and cropping tools, smudge tools, and area selection tools, and provides functions for specifying processing areas and marking error types; Iterative optimization module: Used to execute corresponding optimization strategies based on the specified error type and processing area, and supports multiple rounds of iterative correction; Label output module: Used to save the transparency map finally generated by the iterative optimization module as GT labels and store it in the training set database.

[0019] Reference Figure 1 The embodiments of this application include the following steps: S1: Input the high-resolution original image to be labeled, call at least two pre-stored models for inference, and generate the corresponding number of transparency maps; Specifically, the high-resolution original images to be labeled (resolution ≥ 4K, such as 8K images of 4096×2160 or 7680×4320) enter the system through the data input module. The image data format is RGB three-channel pixel data, and the RGB value of each pixel ranges from 0 to 255. The multi-model inference module pre-stores at least two high-performance image segmentation / matting models such as BiRefNet, InSpyreNet, BEN, and RMBG. These models are all pre-trained and adapted to the local inference requirements of high-resolution images (supporting input size ≥ 512×512).

[0020] The inference process can be performed in parallel or sequentially: In parallel inference, multiple models simultaneously receive the original image data and independently output their respective alphamatte (single-channel data, pixel transparency value range 0-1), which is suitable for scenarios with sufficient computing power; in sequential inference, the model receives the output of the previous model as auxiliary input in turn (such as concatenating the alphamatte of the previous model with the original image and inputting it into the next model), which is suitable for scenarios with limited computing power and can improve inference accuracy.

[0021] S2: Visualize all transparency maps and determine if there are any transparency maps that meet the preset accuracy requirements. Specifically, the visualization filtering module automatically calculates the accuracy index of each alphamatte. If the output result of a certain model meets the preset accuracy requirements, it is directly saved as a GT label (in PNG or TIFF format, retaining complete pixel information) through the label output module for subsequent model training. If the output results of all models do not meet the requirements (such as accuracy index <95% or obvious errors), the interactive modification process is triggered, and the process proceeds to step S3.

[0022] S3: For the error types in the transparency map, optimize by combining the corresponding interactive processing and multi-round iteration strategy; In this process, the interactive operation module provides three core tools: The selection and cropping tool supports both rectangular and polygonal selection modes, with adjustable brush thickness (1-10 pixels). Users drag and select target areas (such as missed main parts or details with insufficient precision) using the mouse. After selection, the system automatically extracts the original pixel data of the area (maintaining high resolution and no compression) and generates a region coordinate file (in JSON format, recording the vertex coordinates of the region (x1, y1), (x2, y2)...(xn, yn), with the origin at the top left corner of the image, the horizontal axis as the x-axis, and the vertical axis as the y-axis). Smudge Tool: Supports round and square brushes with adjustable brush size (5-50 pixels). The preset marker color is pure black (RGB(0,0,0)), but users can customize it to any color with a color difference ≥80% from the subject (difference calculation method: Δ=|R_subject - R_mark|+|G_subject - G_mark|+|B_subject - B_mark|, Δ≥200 is considered a significant difference). If a user misidentifies the marked area as the subject's background, the system records the pixel coordinates of the smudged area in real time and generates a marker mask file (single-channel image, with pixel values ​​of 1 in the marked area and 0 in other areas). Region selection tool: Based on the automatic identification results of low-quality regions, users can directly click on high-priority regions (marked in red) in the heatmap. The system will automatically locate the corresponding position in the original image and generate a recommended selection area. Users can manually adjust the size and shape of the region to reduce the cost of manual inspection.

[0023] The error types include subject recognition errors, foreground omissions, multiple recognitions, and insufficient accuracy in detail areas. When specifying a processing area, the user needs to select the corresponding error type through a drop-down menu. The system binds and stores the area information with the error type, providing a basis for the automatic matching of subsequent optimization strategies. The binding information includes: region coordinates, error type label, specified timestamp, user ID, region size (pixel width × height), and the region's position in the entire image (region pixel count / total image pixel count), ensuring that the problem region is traceable and trackable in each iteration.

[0024] Based on the type of error, the system automatically invokes the corresponding optimization strategy, combining high-resolution local inference, region exclusion, or result fusion to achieve precise correction. The detailed implementation process of each strategy is as follows: A. High-resolution local inference strategy, corresponding to insufficient precision in detailed areas: For detailed areas such as hair strands, semi-transparent fabrics, and glass, inference is performed by maintaining the original resolution of the local area to avoid loss of detailed information. The steps are as follows: Step A1: Based on the coordinates of the detailed region selected by the user, crop the target region from the original high-resolution image (e.g., 4K). The size of the cropped region is in the range of 256×256~2048×2048 (ensure that it meets the minimum input size requirement of the model and does not exceed the GPU memory limit). Extract the RGB three-channel pixel data of the region (RGB value of each pixel is 0-255) without performing any downsampling or compression processing.

[0025] Step A2: Automatically select an image segmentation / matting model suitable for detail segmentation (such as BiRefNet, InSpyreNet, with a focus on optimizing detail region recognition capabilities during pre-training). Set the inference parameters as follows: input size = original size of the cropped region, batch size = 1, inference precision = FP32 (single-precision floating-point), and confidence threshold = 0.9 (only retain pixel results with model prediction confidence ≥ 0.9, and filter out low reliability predictions).

[0026] Step A3: The model performs high-resolution inference on the cropped region and outputs a fine transparency map (alphamatte) of the region. The pixel-level accuracy of the detailed region should reach ≥98% (verified by comparison with manually refined samples). The result includes the transparency value (0-1, 0 for completely transparent and 1 for completely opaque) and the corresponding confidence value (0-1) for each pixel.

[0027] B. Region exclusion strategy, corresponding to multi-recognition: For situations where background objects are misidentified as the main subject, interference is eliminated through marking and masking. The steps are as follows: Step B1: Read the marker mask file generated by the user's brush strokes, identify all marker areas with a pixel value of 1, determine their coordinate range and pixel set, and verify the difference between the marker color and the main color (≥80% is required; if not, prompt the user to brush stroke again).

[0028] Step B2: Concatenate the original image with the marker mask file to form 4-channel data (RGB + marker channel), and input it into the multi-model inference module. During inference, the model masks areas where the marker channel pixel value is 1, forcing the prediction of the transparency value of that area as 0 (completely transparent, judged as background), and only performs foreground subject recognition and transparency calculation on unmarked areas.

[0029] Step B3: After the inference is completed, the system automatically checks whether there are still misjudgments in the marked area (the percentage of pixels with transparency value > 0 is ≤ 0.5% to be considered qualified). If it is not qualified, the system prompts the user to expand the smearing area or adjust the marking color and re-execute the inference process. C. Result fusion strategy, corresponding to subject recognition error and foreground omission: interactively select the target subject range, re-perform multi-model inference for the region (call at least two pre-stored models), select the optimal result, and then fuse it with the uncropped part of the whole image through the weight transparency fusion algorithm, and use the edge smoothing algorithm to process the splicing edge. The specific method is as follows: Step C1: Perform multi-model parallel inference on the user-selected subject area (Crop region) (calling at least two pre-stored models), and calculate the sum of the confidence scores of the output results of each model. The result with the highest total confidence level is selected as the local optimum.

[0030] Step C2: Using a preset weighted transparency fusion algorithm, the local optimal result is fused with the preliminary result of the uncropped portion of the entire image. The algorithm formula and parameters are as follows: ; Where j is the index of the pre-stored model being called. The transparency value of the i-th pixel after merging. Let j be the weight of the j-th model at the i-th pixel. The weight is the transparency value of the i-th pixel output by the j-th model. The calculation method is as follows ,in , is the confidence score of the j-th model at the i-th pixel (range 0-1), output during the model inference process, representing the reliability of the model's transparency prediction result for that pixel.

[0031] During fusion, the model weights corresponding to the local optimum are used first, and the original weight allocation is maintained for the uncropped parts of the whole image.

[0032] Step C3: Edge Smoothing: A bilateral filtering algorithm is used to smooth the merging edge region (the area extending 50 pixels outward from the selected region boundary). The algorithm parameters are set as follows: spatial domain standard deviation. (Control the influence range of pixels within the spatial neighborhood), grayscale standard deviation (Controlling the influence of grayscale similarity on weights) This algorithm adjusts the transparency values ​​of pixels in the edge regions to eliminate stitching marks and ensure the visual coherence and pixel-level consistency of the overall transparency map.

[0033] An iteration can be started if any of the following iteration initiation conditions are met: The initial multi-model inference results did not meet the preset accuracy requirements (e.g., pixel-level accuracy < 95%). After the previous round of optimization, the system detected that there are still unresolved problem areas (the proportion of low-quality areas is greater than the preset threshold, or the user has marked new error areas).

[0034] Step SS1: After initiating the iteration, based on the optimization results of the previous round and combined with the low-quality region identification algorithm, a new round of low-quality region heatmap is generated. Simultaneously, error regions not marked as resolved in the previous round are associated, and a list of regions to be processed in this round is compiled (including region coordinates, error type, previous processing method, and reason for unresolved status). The low-quality region identification algorithm includes: ; in, For the quality estimate of the i-th pixel, Let be the average transparency of n models at the i-th pixel; n is the total number of models involved in the inference. when When T∈[0.03,0.08], the pixel belongs to the low-quality region. This is a preset threshold.

[0035] Step SS2: Display the list of areas to be processed in this round and the heat map on the visualization interface. Users can add new error areas to mark or delete areas that no longer need to be processed (such as mismarked areas) to finally confirm the scope of processing in this round. Step SS3: Based on the error type of the region to be processed in this round, call the corresponding interactive processing and multi-round iteration strategy. The execution process is the same as that of single-round optimization, but only operates on the region to be processed and does not change the results of the regions that have been verified to be correct. Step SS4: After optimization, calculate the accuracy metrics (pixel-level accuracy, percentage of low-quality areas) of this round of optimization results, and generate a verification report to show the number of areas processed in this round, the accuracy comparison before and after optimization, and the status of remaining unresolved areas, so that users can determine whether further iteration is needed.

[0036] The iteration terminates and optimization stops when any of the following conditions are met: The number of iterations has reached the preset limit (users can set 1-5 rounds, default is 3 rounds); The proportion of low-quality areas in the entire map is lower than the preset proportion threshold (users can set 1%-5%, default is 3%). After two consecutive rounds of optimization, if the accuracy improvement is less than 0.5% (accuracy improvement = (accuracy in this round - accuracy in the previous round) / accuracy in the previous round × 100%), the result is considered to have converged. The user can manually terminate the iteration (suitable for rapid annotation needs in special scenarios). S4: Unify and merge the optimized results to generate the final GT labels and save them to the training set.

[0037] After multiple iterations, the system merges the optimization results from all rounds to ensure consistency and high accuracy across the entire graph. Finally, it generates and saves ground truth labels that meet the training requirements. The specific process is as follows: It covers the correct region of the original multi-model inference and the corrected region after optimization in each iteration, ensuring that no region is missed; the correction result of the later iteration has higher priority than the previous iteration, and the local fine inference result has higher priority than the full image fusion result. That is, if the same pixel is corrected multiple times in multiple iterations, the result of the last correction shall prevail; after fusion, the system automatically checks the continuity of pixel values ​​in the full image transparency map (the proportion of pixels with a transparency value difference ≤0.1 between adjacent pixels is ≥99%). If there are abnormal abrupt change regions (the number of consecutive pixels with a difference >0.2 is ≥10), edge smoothing processing is automatically triggered to ensure the rationality of the full image result.

[0038] Generate a single-channel alphamatte with pixel values ​​ranging from 0 to 1 (retaining 6 decimal places to ensure accuracy). Labels are in lossless PNG or TIFF format (supporting 16-bit grayscale storage to avoid accuracy loss due to compression). The label file includes metadata information such as: original image filename, resolution, annotation time, iteration number, model combination used, final pixel-level accuracy, proportion of low-quality regions, and core parameters used in the optimization process (e.g., fusion weight calculation method, edge smoothing algorithm parameters, iteration termination threshold), facilitating subsequent dataset management and traceability.

[0039] Save path rules: Store in a hierarchical structure of "dataset category-image type-annotation date" (e.g., matting dataset / human segmentation / 20260101 / ), file name is consistent with the original image, add "_GT" suffix (e.g., image_001_GT.png).

[0040] Training set format adaptation: Supports exporting to mainstream dataset formats such as COCO and PASCALVOC, automatically generating corresponding annotation information files (such as COCO format JSON annotation files, containing image information, annotation region coordinates, transparency value arrays, etc.), which can be directly used for training image segmentation and matting models without additional format conversion.

[0041] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A high-resolution image segmentation and transparency annotation method, characterized in that, Includes the following steps: S1: Input the high-resolution original image to be labeled, call at least two pre-stored models for inference, and generate the corresponding number of transparency maps; S2: Visualize all transparency images and determine if there is a transparency image that meets the preset accuracy requirements. If so, save it directly as a GT tag; otherwise, proceed to step S3. S3: For the error types in the transparency map, optimize by combining the corresponding interactive processing and multi-round iteration strategy; S4: Unify and merge the optimized results to generate the final GT labels and save them to the training set.

2. The high-resolution image segmentation and transparency annotation method according to claim 1, characterized in that, In step S3, the error types include subject recognition error, foreground omission, multiple recognition, and insufficient accuracy in detail areas; The interaction processing and multi-round iteration strategies corresponding to the aforementioned error types include: Corresponding to subject recognition errors and foreground omissions: interactively select the target subject area, re-perform multi-model inference for that area, select the optimal result, and then fuse it with the uncropped part of the whole image through a weighted transparency fusion algorithm, and use an edge smoothing algorithm to process the stitching edges; Corresponding to multi-recognition: The interaction will paint the misjudged area with the preset background color, and perform multi-model inference again. The model will exclude the background interference area based on the color difference. To address the insufficient precision in detailed areas: interactively crop the detailed areas, perform high-resolution inference on the cropped areas to generate a fine transparency map, and then fuse it with the preliminary results of the full image.

3. The high-resolution image segmentation and transparency annotation method according to claim 2, characterized in that, The method of fusing the image with the uncropped portion of the full image using a weighted transparency fusion algorithm includes: Fusion formula: ; Where j is the index of the pre-stored model being called. The transparency value of the i-th pixel after merging. Let j be the weight of the j-th model at the i-th pixel. The transparency value of the i-th pixel output by the j-th model; During fusion, the model weights corresponding to the local optimum are used first, and the original weight allocation is maintained for the uncropped parts of the whole image.

4. The high-resolution image segmentation and transparency annotation method according to claim 3, characterized in that, The weight The calculation method is as follows: ,in is the confidence score of the j-th model at the i-th pixel, output during the model inference process, representing the reliability of the model's transparency prediction result for that pixel.

5. The high-resolution image segmentation and transparency annotation method according to claim 2, characterized in that, The edge smoothing algorithm uses a bilateral filtering algorithm to smooth the fused edge region.

6. The high-resolution image segmentation and transparency annotation method according to claim 2, characterized in that, The multi-round iteration strategy includes: The iteration is initiated based on the iteration initiation conditions. Based on the optimization results of the previous round and combined with the low-quality region identification algorithm, a new heatmap of low-quality regions is generated. At the same time, the error regions that were not marked as resolved in the previous round are associated with the heatmap and a list of regions to be processed in this round is compiled. The visual interface displays a list of areas to be processed in this round and a heat map, ultimately confirming the scope of processing in this round. Based on the error type of the area to be processed in this round, the corresponding interactive processing and multi-round iteration strategy are invoked, and the execution process is consistent with the single-round optimization. After optimization, the accuracy index of this round of optimization results is calculated, and a verification report is generated, showing the number of regions processed in this round, the accuracy comparison before and after optimization, and the status of remaining unresolved regions, so that users can determine whether further iteration is needed.

7. The high-resolution image segmentation and transparency annotation method according to claim 6, characterized in that, The low-quality region identification algorithm includes: ； in, For the quality estimate of the i-th pixel, Let be the average transparency of n models at the i-th pixel; n is the total number of models involved in the inference. when At that time, the pixel belongs to a low-quality area. This is a preset threshold.

8. The high-resolution image segmentation and transparency annotation method according to claim 1, characterized in that, The pre-stored models include BiRefNet, InSpyreNet, BEN, and RMBG models; The original high-resolution image has a minimum resolution of 4K. During inference, the original resolution of the image is maintained or the original resolution of local areas is maintained, without overall downsampling and compression.

9. A high-resolution image segmentation and transparency annotation system, characterized in that, To implement the method according to any one of claims 1-7, comprising: Data input module: used to receive the high-resolution raw image to be labeled and transmit it to the multi-model inference module; Multi-model inference module: Pre-stores at least two pre-stored models, used to call at least two pre-stored models for inference, and generate a corresponding number of transparency maps; Visualization filtering module: This module is used to visualize all the transparency maps output by the multi-model inference module and provides an accuracy judgment function to filter preliminary results that meet the requirements. Interactive operation module: Provides interactive tools, including selection and cropping tools, smudge tools, and area selection tools, and provides functions for specifying processing areas and marking error types; Iterative optimization module: Used to execute corresponding optimization strategies based on the specified error type and processing area, and supports multiple rounds of iterative correction; Label output module: Used to save the transparency map finally generated by the iterative optimization module as GT labels and store it in the training set database.

Images