Fiber image segmentation method and system based on sparse annotation and dictionary learning

Abstract

The invention relates to a fiber image segmentation method and system based on sparse annotation and dictionary learning. According to the method, small sample training data is quickly constructed by adopting sparse labeling, and probability modeling is performed on local texture features of an image by utilizing dictionary learning, so that high-precision automatic segmentation is realized. In addition, the invention also provides a system for executing the segmentation method and allowing a user to interactively fine-tune the result. Compared with a traditional fully-supervised learning method or an image processing algorithm, the method has the advantages that a high-precision segmentation result can be obtained while the manual annotation cost is greatly reduced, and the dependence on massive training data is reduced. Meanwhile, the segmentation system provided by the invention enables the whole process to be more efficient and accurate, and has scientificity, practicability and wide applicability.

Description

Technical Field

[0001] This invention relates to the field of digital image processing technology, and more particularly to an image segmentation method and system for the microstructure analysis of composite materials. More specifically, it relates to a fiber image segmentation method and system based on sparse annotation and dictionary learning. Background Technology

[0002] Composite materials, due to their high specific strength, high specific modulus, and excellent designability, have been widely used in cutting-edge fields such as aerospace, transportation, and defense industries. The macroscopic mechanical properties of materials are closely related to their internal microstructure, especially the content, distribution, and orientation of reinforcements (such as carbon fibers and glass fibers). Therefore, accurate and efficient quantitative analysis of the microstructure of composite materials is a crucial step in materials research and development, quality control, and failure analysis, and precise fiber segmentation is the foundation for achieving all of this.

[0003] Currently, commonly used fiber image segmentation techniques in the industry can be mainly divided into the following categories:

[0004] (1) Manual segmentation. Technicians typically use general image processing software (such as ImageJ) to select or outline fibers one by one. The advantage of this method is that it is intuitive and has high accuracy in certain situations, but its disadvantages are also very significant: First, it is extremely time-consuming and labor-intensive, and is not suitable for the analysis of large numbers of samples; second, it is highly subjective, and the segmentation results of different operators vary, lacking objectivity and reproducibility; third, the operators need to have certain experience, and the technical threshold is high.

[0005] (2) Traditional automated image processing algorithms. These methods are mainly based on low-level features of images such as grayscale values, gradients, and textures, such as thresholding, edge detection, and watershed algorithms. However, due to the common problems of low contrast, uneven illumination, noise interference, and adhesion and overlap between fibers in microscopic images of composite materials, the robustness of these traditional algorithms is often insufficient. Their segmentation results are extremely sensitive to parameter settings, requiring tedious parameter adjustments for different images, making it difficult to form a universal automated solution.

[0006] (3) Deep learning technology has developed rapidly in recent years. Convolutional neural networks, represented by U-Net, have achieved great success in the field of image segmentation, and can achieve extremely high segmentation accuracy under certain conditions. However, the main bottleneck of this type of method lies in its heavy reliance on a large amount of high-quality labeled data. In order to train an accurate deep learning model, technicians usually need to meticulously depict hundreds or even thousands of images at the pixel level to create a training dataset. This process not only reproduces the tediousness of manual segmentation, but is even more so. The huge time and manpower investment often makes it daunting for many R&D units and specific application scenarios.

[0007] In summary, existing fiber image segmentation techniques generally face a dilemma in balancing manual costs, algorithm robustness, and data dependence. Therefore, there is an urgent need in this field for a novel fiber image segmentation technique that can balance efficiency, accuracy, and practicality. This technique should be able to adapt quickly to different image conditions with minimal manual annotation costs and achieve robust and reliable automated segmentation, thereby addressing the shortcomings of the existing technologies. Summary of the Invention

[0008] The purpose of this invention is to provide a fiber image segmentation method and system based on sparse annotation and dictionary learning, which can significantly reduce labor costs and improve the adaptability of segmentation results to different material samples and imaging conditions, thereby solving the problems mentioned in the background art.

[0009] To achieve the above objectives, the first technical solution adopted by this invention is: a fiber image segmentation method based on sparse labeling and dictionary learning, comprising the following steps:

[0010] Step S1: Obtain a sample image and receive sparse annotation information from the user on the sample image for at least two preset categories (e.g., fiber and background).

[0011] Step S2: Based on the sparse annotation information, extract the first image patch set only from the marked pixel positions in the sample image;

[0012] Step S3: Perform cluster analysis on the first image patch set to generate a texture dictionary containing multiple dictionary atoms, where each dictionary atom represents a specific local image texture;

[0013] Step S4: Based on the sparse annotation information, the probability of each dictionary atom in the texture dictionary belonging to each preset category is statistically calculated, thereby establishing a probability mapping relationship between dictionary atoms and preset categories to obtain a probabilistic texture dictionary;

[0014] Step S5: Obtain an image to be segmented, and densely extract a second set of image patches from it;

[0015] Step S6: For each image block in the second image block set, find the dictionary atom that best matches it in the probabilistic texture dictionary, and determine the probability that the image block belongs to each preset category according to the probability mapping relationship, thereby generating a pixel-level category probability map of the image to be segmented.

[0016] Step S7: Based on the probability map, the segmentation of the image to be segmented is completed by selecting the category with the highest probability at each pixel location as its final category.

[0017] Furthermore, as a preferred embodiment, before step S1, the method of the present invention further includes a preprocessing step: receiving physical parameters related to the image (e.g., the actual diameter of the fiber and the pixel size of the image), and performing adaptive scaling on the image based on the physical parameters, so that the target features in the image are normalized to a preset range in pixel size.

[0018] Furthermore, as a preferred embodiment, after step S7, the method of the present invention further includes an interactive fine-tuning step: presenting the pixel-level category probability map to the user and providing an interactive tool that allows the user to adjust the probability threshold to correct and confirm the final segmentation result in real time.

[0019] The second technical solution adopted in this invention is: a fiber image segmentation system based on sparse annotation and dictionary learning, comprising the following components:

[0020] Input module: used to receive initial instruction information input by the user, the initial instruction information including at least physical parameters and selection instructions for the region of interest;

[0021] Preprocessing module: used to perform adaptive scaling on the image based on the physical parameters received by the input module;

[0022] Annotation module: Used to provide an interactive graphical interface and receive sparse annotation information of sample images by users on the interface;

[0023] Model training module: used to construct and generate a probabilistic texture dictionary based on the sparse annotation information received by the annotation module;

[0024] Image segmentation module: Used to process the target image using the probabilistic texture dictionary generated by the model training module, and output the segmentation result;

[0025] Data access module: used to read image data from external storage devices, and to save or read the probabilistic texture dictionary generated by the model training module;

[0026] It also includes a processor and a memory, wherein the memory contains computer program instructions that enable the processor to perform the functions of the aforementioned modules.

[0027] The beneficial effects of the method and system of this invention are as follows: Using sparse annotation instead of pixel-level full annotation, users only need to perform a small amount of masking on representative areas to complete the preparation of training samples, significantly reducing manual costs and improving efficiency. Since the texture dictionary is constructed entirely based on user-specified regions with clear semantics, the model only learns key textures for the target task, effectively eliminating interference from irrelevant background information. The model is highly targeted and has high segmentation accuracy. By introducing physical parameters for adaptive scale normalization preprocessing, this invention can effectively handle images with different magnifications and resolutions, making a single model applicable to a wider variety of image sources. The algorithm is robust and has good universality. An interactive fine-tuning function based on probabilistic graphs is provided, transforming the model's "uncertainty" into intuitive adjustment criteria, allowing users to quickly optimize the segmentation results. This human-machine collaborative closed-loop control significantly improves the system's practicality and the reliability of the final results. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0029] Figure 1 A fiber image segmentation system based on sparse labeling and dictionary learning.

[0030] Figure 2 Flowchart of a fiber image segmentation method based on sparse labeling and dictionary learning

[0031] Figure 3 Schematic diagram of the Region of Interest (ROI) selection interface

[0032] Figure 4 Sparse annotation interface diagram

[0033] Figure 5 Comparison diagram of segmentation results under high resolution conditions

[0034] Figure 6 Comparison diagram of segmentation results under low resolution conditions

[0035] Figure 7 Interactive fine-tuning interface diagram Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] Reference Figure 1A fiber image segmentation system based on sparse labeling and dictionary learning comprises:

[0038] Input module: Used to receive initial instructions from the user, such as selecting an image file, inputting physical parameters, and selecting a region of interest;

[0039] Preprocessing module: Used to perform adaptive scaling on the image based on the physical parameters received by the input module;

[0040] The annotation module provides an interactive graphical interface for users to perform sparse annotations on sample images and receive the resulting annotation information.

[0041] Model training module: Used to execute a dictionary learning algorithm based on the sparse annotation information received by the annotation module, in order to construct and generate a probabilistic texture dictionary;

[0042] Image segmentation module: Used to automatically segment target images using the probabilistic texture dictionary generated by the model training module, and output the segmentation results;

[0043] Data access module: Responsible for data interaction with external storage devices, including reading raw image data and saving or loading trained models.

[0044] Reference Figure 2 This paper demonstrates a fiber image segmentation method based on sparse annotation and dictionary learning. The following section will combine... Figure 1 as well as Figures 3 to 7 The process of this method is described in detail, and the process corresponds to steps S1 to S7 of the method claimed in this invention.

[0045] In a preferred embodiment of the present invention, the system may perform a preprocessing step before the method begins. The user inputs physical parameters related to the image via an input module, such as the actual diameter of the fiber and the pixel size of the image. Subsequently, the preprocessing module automatically calculates a scaling factor based on these parameters and adaptively scales the image so that the target features in the image are normalized to a preset range in pixel size. This step can effectively handle images of different resolutions, such as... Figure 5 (High resolution) and Figure 6 As shown in (low resolution), this greatly improves the adaptability of the method.

[0046] Step S1: Obtain a sample image and receive sparse annotation information from the user on the sample image for at least two preset categories.

[0047] Specifically, the user loads a sample image through the data access module and selects a region of interest (ROI) on that image through the input module, such as... Figure 3 As shown. Then, as... Figure 4 As shown, the annotation module provides an interactive interface where users can perform sparse annotations, such as using the first type of marker (representing the background class) and the second type of marker (representing the fiber class) to make small amounts of smears in the corresponding areas. The annotation module receives and records these sparse annotation information.

[0048] Step S2: Based on the sparse annotation information, extract the first image patch set only from the marked pixel positions in the sample image.

[0049] This step is performed by the model training module. This module reads the annotation information recorded by the annotation module in step S1, and extracts a series of fixed-size local image patches only at the marked pixel coordinates. These image patches constitute the first image patch set.

[0050] Step S3: Perform cluster analysis on the first image patch set to generate a texture dictionary containing multiple dictionary atoms.

[0051] The process continues with the model training module. This module applies a clustering algorithm (K-Means algorithm in this embodiment) to the first set of image patches obtained in step S2, grouping image patches with similar texture features together. The center vector of each cluster is defined as a "dictionary atom," representing a specific local image texture. All these dictionary atoms together constitute the texture dictionary.

[0052] Step S4: Based on the sparse annotation information, establish a probability mapping relationship between dictionary atoms and preset categories to obtain a probabilistic texture dictionary.

[0053] This step is still performed by the model training module. This module backtracks the original class labels (from the sparse annotations in step S1) of each image patch in the first image patch set and counts the number of image patches from the "background" and "fiber" classes contained in each dictionary atom (i.e., each cluster). Based on this statistical result, a probability distribution is calculated for each dictionary atom, thus obtaining a probabilistic texture dictionary containing the probability information of each texture atom belonging to each class. This model can be saved through the data access module.

[0054] Step S5: Obtain an image to be segmented and densely extract a second set of image patches from it.

[0055] When a new image needs to be segmented, the user selects the image or ROI to be segmented via the input module. The image segmentation module receives the image and extracts dense local image patches of the same size as during training by sliding pixel by pixel. These image patches constitute the second set of image patches.

[0056] Step S6: For each image block in the second image block set, determine the probability that it belongs to each preset category, thereby generating a pixel-level category probability map of the image to be segmented.

[0057] This step is performed by the image segmentation module. For each image patch in the second image patch set, the module searches for the best-matching dictionary atom (e.g., closest Euclidean distance) in the probabilistic texture dictionary generated in step S4, and assigns the class probability corresponding to that dictionary atom to the image patch. After traversing all image patches, the system calculates the probability of each pixel in the image to be segmented belonging to each class, thereby generating a pixel-level "target class probability map," such as... Figure 7 As shown.

[0058] Step S7: Based on the probability map, the segmentation of the image to be segmented is completed by selecting the category with the highest probability at each pixel location as its final category.

[0059] The image segmentation module, based on the probability map generated in step S6, selects the category with the highest probability value for each pixel as its final segmentation category, thereby generating the final segmentation result map. For example... Figure 5 and Figure 6 As shown, this method can achieve accurate segmentation results for images under different imaging conditions.

[0060] In a preferred embodiment of the present invention, after step S7, the system further provides an interactive fine-tuning step. For example... Figure 7 As shown, the image segmentation module presents the target category probability map to the user. The user can adjust the probability threshold through an input module. The system updates the final segmentation result based on the new threshold set by the user, allowing for intuitive and quick correction and confirmation of the results.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fiber image segmentation method based on sparse labeling and dictionary learning, characterized in that, Includes the following steps: Step S1: Obtain a sample image and receive sparse annotation information from the user on the sample image for at least two preset categories (e.g., fiber and background). Step S2: Based on the sparse annotation information, extract the first image patch set only from the marked pixel positions in the sample image; Step S3: Perform cluster analysis on the first image patch set to generate a texture dictionary containing multiple dictionary atoms, where each dictionary atom represents a specific local image texture; Step S4: Based on the sparse annotation information, the probability of each dictionary atom in the texture dictionary belonging to each preset category is statistically calculated, thereby establishing a probability mapping relationship between dictionary atoms and preset categories to obtain a probabilistic texture dictionary; Step S5: Obtain an image to be segmented, and densely extract a second set of image patches from it; Step S6: For each image block in the second image block set, find the dictionary atom that best matches it in the probabilistic texture dictionary, and determine the probability that the image block belongs to each preset category according to the probability mapping relationship, thereby generating a pixel-level category probability map of the image to be segmented. Step S7: Based on the probability map, the segmentation of the image to be segmented is completed by selecting the category with the highest probability at each pixel location as its final category.

2. The method according to claim 1, characterized in that, Before step S1, a preprocessing step is included: receiving physical parameters related to the image and performing adaptive scaling on the sample image and the image to be segmented based on the physical parameters.

3. The method according to claim 2, characterized in that, The physical parameters include at least the actual diameter of the fiber and the pixel size of the image, and the adaptive scaling normalizes the fiber features in the image to a preset range in terms of pixel size.

4. The method according to claim 1, characterized in that, Following step S7, an interactive fine-tuning step is also included: presenting the pixel-level category probability map to the user and providing an interactive tool that allows the user to adjust the probability threshold to correct and confirm the final segmentation result.

5. The method according to claim 1, characterized in that, The clustering algorithm used in step S3 is the K-Means algorithm.

6. The method according to claim 1, characterized in that, In step S6, finding the most similar dictionary atom is achieved by calculating the Euclidean distance and selecting the dictionary atom with the closest distance.

7. A fiber image segmentation system based on sparse labeling and dictionary learning, characterized in that, include: Input module: used to receive initial instruction information input by the user, the initial instruction information including at least physical parameters and selection instructions for the region of interest; Preprocessing module: used to perform adaptive scaling on the image based on the physical parameters received by the input module; Annotation module: Used to provide an interactive graphical interface and receive sparse annotation information of sample images by users on the interface; Model training module: used to execute steps S2 to S4 of claim 1 based on the sparse annotation information received by the annotation module, and to construct and generate a probabilistic texture dictionary; Image segmentation module: used to execute steps S5 to S7 of claim 1, to process the target image using the probabilistic texture dictionary generated by the model training module, and output the segmentation result; Data access module: used to read image data from external storage devices, and to save or read the probabilistic texture dictionary generated by the model training module; It also includes a processor and a memory, the memory containing computer program instructions which, when executed by the processor, are used to implement the functions of the aforementioned modules.

8. The system according to claim 7, characterized in that, The image segmentation module is also used to: present the segmentation results and pixel-level category probability maps to the user, receive instructions from the user to adjust the probability threshold, and then update the segmentation results in real time according to the instructions to achieve interactive fine-tuning.

9. The system according to claim 7, characterized in that, The clustering algorithm used in the model training module is the K-Means algorithm.

10. The system according to claim 7, characterized in that, The image segmentation module finds the most similar dictionary atom by calculating the Euclidean distance and selecting the closest dictionary atom.

Images