A method and system for membrane structure damage detection and identification that integrates depth features and traditional texture analysis
By integrating deep learning with traditional texture analysis, and utilizing an improved EAC-YOLO network and Gabor filter, combined with the SSIM differential enhancement algorithm, the problems of texture interference and missed detection of minute defects in membrane structure damage detection are solved, achieving high-precision and convenient damage identification and analysis.
Patent Information
- Application Number
- CN202611415319.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-13
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for membrane structure damage detection suffer from problems such as large texture interference, high rate of missed detection of minute defects, limited detection methods, and inconvenient human-computer interaction, making it difficult to achieve high-precision automatic identification and rapid texture damage analysis.
This invention integrates an improved EAC-YOLO deep network with traditional detection algorithms that combine multi-texture mode Gabor suppression and SSIM difference enhancement. It enables intelligent identification and detailed analysis of membrane structure damage through a graphical interface, including texture suppression, local structural similarity calculation, difference map enhancement, and damage mask generation.
It achieves accurate segmentation and identification of membrane structure damage against complex texture backgrounds, improves the detection rate of minute damage, provides a convenient human-computer interaction interface, and meets the needs of high-precision automatic identification and rapid texture damage analysis under sample-less conditions.
Smart Images

Figure CN123118581A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to a method and system for membrane structure damage detection and recognition that integrates depth features and traditional texture analysis, particularly a method and system for membrane structure damage detection and recognition that integrates deep learning target detection, multi-modal texture suppression, and structural similarity anomaly analysis. Background Technology
[0002] Membrane structure materials (such as PTFE and PVC membranes) are widely used in large-span buildings such as stadiums and transportation hubs due to their light weight and good light transmittance. However, under wind loads, stress aging, and environmental erosion, membrane materials are prone to typical damage such as cracks, holes, wrinkles, and stains. Current vision-based inspection methods have the following main shortcomings:
[0003] 1. Significant background interference: The membrane surface has complex woven textures such as oblique, orthogonal and sinusoidal curves, which, combined with uneven lighting, make it difficult for traditional image processing algorithms to accurately segment the damaged area.
[0004] 2. Difficulty in detecting small targets: Early cracks account for a very small percentage of pixels in an image, and general deep learning models have low recall rates for such minute damages.
[0005] 3. Disconnect between algorithms and business: Existing detection methods either only provide deep learning inference or only provide traditional filtering and segmentation, lacking an intuitive operating interface that organically integrates the two and is geared towards field engineers. This makes it difficult to simultaneously meet the needs of high-precision automatic identification and rapid texture damage analysis under sample-less conditions. Summary of the Invention
[0006] Therefore, the purpose of this application is to provide a method and system for membrane structure damage detection and identification that integrates depth features and traditional texture analysis, in order to solve the technical problems in the prior art, such as damage segmentation being greatly affected by texture interference, high rate of missed detection of small defects, limited detection methods, and inconvenient human-computer interaction. In particular, this invention integrates an improved EAC-YOLO deep network with a traditional detection algorithm that combines multi-texture mode Gabor suppression and SSIM difference enhancement into a single graphical interface, thereby achieving intelligent identification and refined analysis of membrane structure damage.
[0007] To achieve the above objectives, this application provides the following technical solution:
[0008] A method for membrane structure damage detection and identification that integrates depth features and traditional texture analysis includes:
[0009] Acquire images of the membrane structure to be detected;
[0010] A first detection mode and a second detection mode are provided for users to choose from;
[0011] In the first detection mode, the trained improved EAC-YOLO network is used to perform damage recognition on the image, outputting damage category, confidence level and bounding box coordinates, and the detection results are saved as a text file, which contains the category, confidence level and normalized bounding box coordinate information of each detected target;
[0012] In the second detection mode, perform the following steps:
[0013] Step 1: Based on the membrane texture type selected by the user, call the corresponding improved Gabor filter parameter group to suppress the background texture of the image; the texture type includes oblique texture, orthogonal texture, and sinusoidal curve texture;
[0014] Step 2: Perform Local Structural Similarity (SSIM) calculation on the texture-suppressed image and the pre-stored non-destructive reference image to obtain the SSIM difference map. Then, perform gamma correction, contrast-limited adaptive histogram equalization (CLAHE), and Laplacian edge enhancement on the difference map in sequence. After weighted fusion, obtain the enhanced difference map.
[0015] Step 3: Perform adaptive threshold segmentation on the enhanced difference map to obtain an initial damage mask; calculate the texture energy difference map to generate an auxiliary mask; fuse the initial damage mask with the auxiliary mask; and obtain the final damage mask after morphological optimization.
[0016] Step 4: Perform connected component analysis based on the final damage mask, extract the damage region, generate a damage matting map with a white background, and output the area, perimeter, circularity, and aspect ratio features of the damage region.
[0017] Preferably, the improved Gabor filter parameter set in step 1 includes: filters with directions of 0° and 90° and center frequencies of 0.1 to 0.2 for oblique textures; filters with directions of 45° and 135° and center frequencies of 0.1 to 0.2 for orthogonal textures; and filters with multiple directions and direction intervals of no more than 30° and center frequencies of 0.1 to 0.2 for sinusoidal textures. The Gaussian envelope scale of each parameter set is adjusted according to the period of the corresponding texture.
[0018] Preferably, the enhancement of the SSIM difference map in step 2 specifically involves: normalizing the difference map and then performing a gamma transform to enhance high-value regions, then using CLAHE to improve local contrast, while extracting the edges of the difference map using the Laplacian operator, and finally weighting and fusing the CLAHE-enhanced map and the Laplacian edge map with a weight of 0.6:0.4.
[0019] Preferably, the adaptive threshold segmentation in step 3 employs the Otsu method or the adaptive Gaussian threshold method; the morphological optimization includes performing closing and opening operations on the mask using elliptical structuring elements to remove noise and connect broken and damaged regions.
[0020] Preferably, in the second detection mode, the method further includes saving the damage cutout, damage marker map, and features to a designated folder, and synchronously displaying each intermediate processing result in the graphical user interface.
[0021] This invention also provides a system for membrane structure damage detection and identification that integrates depth features and traditional texture analysis, comprising:
[0022] A first detection module, comprising:
[0023] (1) EAC-YOLO network inference unit, used to load pre-trained weights or user-specified weight files, perform damage detection on input images, video streams or real-time camera footage, and output damage category, confidence level and location coordinates;
[0024] (2) Result export unit, used to save the image detection results as a text file, wherein the text file contains the category, confidence level and normalized bounding box coordinates of each detected target;
[0025] The second detection module includes:
[0026] (1) Image preprocessing unit, used to perform grayscale conversion, Canny edge detection and binarization operations on the input image;
[0027] (2) Texture suppression unit, which stores multiple sets of improved Gabor filter parameters corresponding to oblique texture, orthogonal texture and sinusoidal texture respectively, and is used to suppress background texture of the image according to the texture type selected by the user;
[0028] (3) SSIM damage detection unit, used to perform local SSIM analysis on the texture-suppressed image and the reference image, and generate a damage mask through difference map enhancement, threshold segmentation and energy map fusion;
[0029] (4) Damage extraction unit, used to extract the damage area according to the damage mask, generate a damage matting with a white background, and calculate the area, perimeter, roundness and aspect ratio of the damage area;
[0030] (5) A graphical user interface, which is connected to the first detection module and the second detection module respectively, and provides detection mode switching, parameter setting, result display and result export functions.
[0031] Preferably, the improved Gabor filter parameter group in the texture suppression unit is specifically: a filter group corresponding to the oblique texture in the direction [0°, 90°]; a filter group corresponding to the orthogonal texture in the direction [45°, 135°]; and a filter group corresponding to the sinusoidal texture in the direction [0°, 30°, 60°, 90°, 120°, 150°], and the center frequency and scale of each filter group are preset according to the dominant frequency of the corresponding texture.
[0032] Preferably, the SSIM damage detection unit generates the damage mask in the following manner: after calculating the SSIM difference map, it is sequentially subjected to gamma correction, CLAHE and Laplacian edge enhancement and weighted fusion to obtain an enhanced difference map; the enhanced difference map is subjected to adaptive threshold segmentation to obtain an initial mask; the texture energy difference map between the image to be tested and the reference image is calculated and binarized as an auxiliary mask; the initial mask and the auxiliary mask are fused, and then optimized by morphological closing and opening operations and inverted to obtain a final mask with a white background and black damage.
[0033] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the method according to any one of claims 1 to 5.
[0034] The system of this invention comprises two main functional modules: the first module is a membrane structure surface damage identification module based on EAC-YOLO; the second module is a damage image processing module based on OpenCV. The two modules share an image acquisition interface and use a unified PyQt5 graphical user interface for mode switching and result display.
[0035] Module 1: Damage Recognition Module Based on EAC-YOLO
[0036] This module is used for deep learning detection of various types of damage on membrane structures, including cracks, pores, wrinkles, and stains. Its core is an improved YOLO network, EAC-YOLO, which introduces an ECA attention mechanism into the backbone network to enhance sensitivity to subtle defects and textures, employs the ADown downsampling operator to reduce computation, and utilizes CGBlock to reconstruct the C3k2 module to improve feature discrimination capabilities. The module supports loading pre-trained EAC-YOLO weight files or other weight files trained by the user using their own datasets. Recognized patterns include:
[0037] 1. Image recognition: After uploading the image, inference and annotation are performed. The detection results can be exported as a txt file. The file contains the damage category, confidence level and normalized bounding box coordinates of each detected target (format: category confidence level x_center y_center width height).
[0038] 2. Folder recognition: Traverse the images within a folder, process them asynchronously using multiple threads, and display the progress.
[0039] 3. File recognition: Read local MP4 video streams, extract frames for inference, and render in real time.
[0040] 4. Detection: Capture images via RTSP or USB camera and perform online detection at the set frame rate.
[0041] The test results are displayed in both table and annotated graph formats on the interface. Clicking on a table row will automatically zoom in on the corresponding damaged area.
[0042] Module 2: Damage Image Processing Module Based on OpenCV
[0043] This module provides basic image segmentation functions such as grayscale conversion, Canny edge detection, and binarization, and introduces an improved damage handling algorithm, which includes the following three processing stages:
[0044] (1) Improved Gabor texture suppression: Based on the common woven texture types on the surface of membrane structures, textures are divided into three categories: oblique textures, orthogonal textures, and sinusoidal textures. The system has three sets of optimized Gabor filter parameters. Users can select the corresponding texture mode according to the membrane material under test to perform background texture directional suppression. Each set of parameters is specifically designed in terms of filter direction, center frequency, and Gaussian envelope scale to significantly reduce the response of regular textures while protecting the damaged edges.
[0045] (2) Damage detection based on SSIM difference enhancement: Local structural similarity (SSIM) calculation is performed on the texture-suppressed test image and the pre-acquired undamaged reference image to obtain a difference map. Gamma correction, contrast-limited adaptive histogram equalization (CLAHE), and Laplacian edge enhancement are sequentially applied to the difference map, and then fused according to weights to form an enhanced difference map. Subsequently, an adaptive threshold (Otsu's method or Gaussian adaptive) is used to segment the enhanced difference map to obtain an initial damage mask. Simultaneously, the texture energy difference map of the two images is calculated and used as an auxiliary mask, which is fused with the initial mask. After morphological closing and opening operations optimization, the final damage mask is generated. In this mask, the damaged area is black and the background is white, achieving robust separation of damage against complex texture backgrounds.
[0046] (3) Damage region extraction and feature analysis: Connectivity analysis is performed on the final damage mask to accurately extract each damage region on a white background while preserving its original grayscale features. The algorithm automatically calculates and outputs the area, perimeter, circularity, aspect ratio, average grayscale and position coordinates of each damage region, and saves the individual damage cutout and contour marking map to the specified directory for subsequent quantitative evaluation.
[0047] Graphical User Interface: The two modules described above are integrated into the same PyQt5 interface. The interface includes a mode switching tab, weight loading and selection controls, a texture type dropdown menu, a detection trigger button, an image preprocessing and result display canvas, a damage list table, and a result export button. A frame rate adjustment slider is also provided in real-time detection mode.
[0048] The membrane structure damage detection and identification method and system that integrates depth features and traditional texture analysis provided in this application has the following advantages:
[0049] 1. Complementary detection methods: The deep learning module can accurately identify multiple types of damage and output structured detection results; the traditional image processing module can quickly locate damage through texture suppression and SSIM anomaly detection without training samples, meeting different field conditions.
[0050] 2. Strong texture targeting: The improved Gabor filter provides three texture suppression modes: oblique, orthogonal, and quasi-sine curve, which effectively overcomes the interference of complex background of membrane structure on damage segmentation and significantly improves the purity of damage mask.
[0051] 3. Refined damage analysis: The multi-level enhancement strategy of SSIM difference map and energy map fusion improves the detection rate of micro-cracks and low-contrast pores; the damage extraction stage not only provides image matting, but also outputs geometric and grayscale features to assist in damage level assessment.
[0052] 4. Convenient and efficient interaction: The integrated UI supports various workflows from real-time monitoring to batch post-processing. The test results can be directly exported as txt files or saved as images, reducing the technical threshold for membrane structure maintenance. Attached Figure Description
[0053] Figure 1 This is a diagram of the overall system architecture of the present invention.
[0054] Figure 2 This is a screenshot of the system homepage interface of the present invention.
[0055] Figure 3 This is a schematic diagram of the EAC-YOLO network structure in this invention.
[0056] Figure 4 This is a layout diagram of the graphical user interface in this invention.
[0057] Figure 5 This is a flowchart of the improved Gabor filtering and SSIM damage detection algorithm in this invention.
[0058] Figure 6 This is the original image of the crack damage and the identification results in this invention.
[0059] Figure 7The image shows the original diagram and identification results of the hole damage in this invention.
[0060] Figure 8 This is the original image and identification result of the scratch damage in this invention.
[0061] Figure 9 This is the original image of the stain damage and the identification result in this invention.
[0062] Figure 10 This is the original image and identification result of the wrinkle damage in this invention.
[0063] Figure 11 This is the original image and identification result of multiple damages in this invention. Detailed Implementation
[0064] like Figure 1-11 As shown, a method for membrane structure damage detection and identification that integrates depth features and traditional texture analysis includes:
[0065] Acquire images of the membrane structure to be detected;
[0066] A first detection mode and a second detection mode are provided for users to choose from;
[0067] In the first detection mode, the trained improved EAC-YOLO network is used to perform damage recognition on the image, outputting damage category, confidence level and bounding box coordinates, and the detection results are saved as a text file, which contains the category, confidence level and normalized bounding box coordinate information of each detected target;
[0068] In the second detection mode, perform the following steps:
[0069] Step 1: Based on the membrane texture type selected by the user, call the corresponding improved Gabor filter parameter group to suppress the background texture of the image; the texture type includes oblique texture, orthogonal texture, and sinusoidal curve texture;
[0070] Step 2: Perform Local Structural Similarity (SSIM) calculation on the texture-suppressed image and the pre-stored non-destructive reference image to obtain the SSIM difference map. Then, perform gamma correction, contrast-limited adaptive histogram equalization (CLAHE), and Laplacian edge enhancement on the difference map in sequence. After weighted fusion, obtain the enhanced difference map.
[0071] Step 3: Perform adaptive threshold segmentation on the enhanced difference map to obtain an initial damage mask; calculate the texture energy difference map to generate an auxiliary mask; fuse the initial damage mask with the auxiliary mask; and obtain the final damage mask after morphological optimization.
[0072] Step 4: Perform connected component analysis based on the final damage mask, extract the damage region, generate a damage matting map with a white background, and output the area, perimeter, circularity, and aspect ratio features of the damage region.
[0073] Preferably, the improved Gabor filter parameter set in step 1 includes: filters with directions of 0° and 90° and center frequencies of 0.1 to 0.2 for oblique textures; filters with directions of 45° and 135° and center frequencies of 0.1 to 0.2 for orthogonal textures; and filters with multiple directions and direction intervals of no more than 30° and center frequencies of 0.1 to 0.2 for sinusoidal textures. The Gaussian envelope scale of each parameter set is adjusted according to the period of the corresponding texture.
[0074] Preferably, the enhancement of the SSIM difference map in step 2 specifically involves: normalizing the difference map and then performing a gamma transform to enhance high-value regions, then using CLAHE to improve local contrast, while extracting the edges of the difference map using the Laplacian operator, and finally weighting and fusing the CLAHE-enhanced map and the Laplacian edge map with a weight of 0.6:0.4.
[0075] Preferably, the adaptive threshold segmentation in step 3 employs the Otsu method or the adaptive Gaussian threshold method; the morphological optimization includes performing closing and opening operations on the mask using elliptical structuring elements to remove noise and connect broken and damaged regions.
[0076] Preferably, in the second detection mode, the method further includes saving the damage cutout, damage marker map, and features to a designated folder, and synchronously displaying each intermediate processing result in the graphical user interface.
[0077] Example 1: Workflow of EAC-YOLO Damage Recognition Module
[0078] Step 1: Training and Weight Generation
[0079] The optimal .pt weight file was generated by training on five labeled datasets of membrane structure damage. In this experiment, Windows 11 was used. The GPU used was an NVIDIA GeForce RTX 4070 Super with 12GB of VRAM; the CPU was an Intel(R) Core(TM) i9-14900K 3.2 GHz; Python version was 3.9.20; the deep learning framework and version used were PyTorch 2.0.0 and CUDA version 11.8.
[0080] After training, the training results of EAC-YOLO on the five-class damage test set of membrane structures are shown in Table 1. Precision (P), recall (R), number of parameters, mean average precision (mAP@0.5), and computational cost (GFLOPs) are used as evaluation metrics for the model. Precision (P) refers to the proportion of correctly detected positive samples out of all predicted positive samples; recall (R) is the proportion of correctly detected positive samples out of all true positive samples; mAP@0.5 refers to the mean average precision of all classes in the detection task when the IoU threshold is 0.5, reflecting the model's generalization ability in multi-class scenarios; mAP@0.5:0.95 measures the model's average precision within the IoU threshold range of 0.5 to 0.95; FLOPs is the number of floating-point operations performed by the model, used to assess computational resource requirements and measure the complexity of the algorithm model. Related training results are shown in Table 2.
[0081] Table 1. EAC-YOLO Training Related Parameters
[0082]
[0083] Table 2 EAC-YOLO Training Results
[0084]
[0085] As shown in Table 2, EAC-YOLO achieved a high average accuracy on the five types of membrane structure damage datasets, especially for the identification of holes, stains and cracks. It also maintained usable detection accuracy for damages with complex deformation and blurred boundaries, such as wrinkles and coating peeling, which verified the effectiveness of ECA attention mechanism, ADown downsampling and CGBlock reconstruction for extracting membrane structure damage features.
[0086] Step 2: The user selects the "Damage Recognition" mode through the graphical interface, clicks the "Load Weights" button to select a pre-trained weight file (.pt), or directly uses the system default EAC-YOLO.pt. These weights are the EAC-YOLO network parameters trained on a dataset of five types of membrane structure damage.
[0087] Step 3: Select the recognition mode "Image Recognition" and upload the image of the membrane structure surface to be detected. The system calls OpenCV to read the image, scales it to 640×640, and inputs it into the EAC-YOLO network. The network extracts multi-scale damage features through ECA attention-enhanced feature extraction, ADown downsampling, and the reconstructed C3k2 module, and outputs predicted bounding boxes.
[0088] Step 4: Draw bounding boxes with category labels on the results canvas. The table on the right simultaneously lists "Type", "Confidence", and "Location Coordinates". When the user clicks "Export Results", the system writes the detection results of the current image to a txt file. Each line is formatted as: "Crack 0.92 0.453 0.521 0.083 0.106", representing the category, confidence, normalized center coordinates, and width and height, respectively.
[0089] Step 5: If "Real-time Detection" is selected, the system will start the USB camera, extract frames at 30 FPS, continuously infer and refresh the screen to achieve online monitoring of the membrane structure on site.
[0090] Furthermore, the recognition accuracy of EAC-YOLO in Example 1 on five types of membrane structure damage test sets was statistically analyzed, and the statistical results are shown in Table 3.
[0091] Table 3 Summary of EAC-YOLO Damage Recognition Experiment Results in Example 1
[0092]
[0093] As shown in Table 3, EAC-YOLO in Example 1 achieved high accuracy rates in identifying pores, stains, and cracks, reaching 95.0%, 93.6%, and 92.5%, respectively; and accuracy rates in identifying wrinkles and coating peeling, reaching 88.9% and 90.0%, respectively. The overall accuracy rate for identifying the five types of damage was 92.2%, meeting the requirements for automatic identification of surface damage in membrane structures. Figure 6 This is a comparison chart of typical EAC-YOLO recognition results for the five types of damage shown in Table 2.
[0094] The judgment criteria are explained below:
[0095] Correct identification: The damage category output by EAC-YOLO is consistent with the manual annotation, and the intersection-union ratio (IoU) between the detection box and the manual annotation box is ≥0.5;
[0096] False positive: Undamaged areas were marked as damaged by EAC-YOLO;
[0097] False negative: Damaged areas contained in the sample were not detected by EAC-YOLO;
[0098] The average detection time is the average inference time for a single 640×640 image on an NVIDIA RTX 4070 Super platform.
[0099] Example 2: Workflow of Damage Handling Module Based on Improved Gabor and SSIM
[0100] Step 1: The user switches to "Traditional Image Processing" mode and loads the surface image of the membrane structure to be detected. In the preprocessing area, grayscale conversion and Canny edge detection can be used to observe the damage contour.
[0101] Step 2: Depending on the membrane material type, select "Oblique Texture," "Orthogonal Texture," or "Sine-like Texture" from the "Texture Suppression" drop-down menu. The system calls the corresponding Gabor filter banks: oblique texture uses directions [0°, 90°] and a center frequency of 0.15; orthogonal texture uses [45°, 135°] and a center frequency of 0.15; sine-like texture uses multiple directions [0°, 30°, 60°, 90°, 120°, 150°] with a center frequency of 0.15, and the corresponding scale parameter σ. Perform a two-dimensional convolution on the image to obtain the texture-suppressed image.
[0102] Step 3: Load a non-destructive reference image of the same scene (either preset or acquired on-site). Calculate the local SSIM maps of the texture-suppressed image and the reference image to obtain a difference map. After normalizing the difference map, perform a gamma transform (γ=1.5) to enhance high-discrepancy areas, then use CLAHE to improve contrast, and finally use the Laplacian operator to extract edges. Fuse the CLAHE-enhanced image and the Laplacian edge map with weights of 0.6 and 0.4 respectively to form an enhanced difference map.
[0103] Step 4: Use Otsu's method to obtain an initial damage mask (damage is white) on the enhanced difference map; at the same time, calculate the difference in texture energy maps between the two maps and obtain an auxiliary mask using Otsu's method; after fusing the two masks, perform elliptic kernel morphological closing and opening operations to obtain the final damage mask, and invert it to black damage and white background.
[0104] Step 5: Based on the final mask, perform connected component labeling and extract each damage region. Generate a damage matting map with a white background, calculate features such as area, perimeter, circularity, and aspect ratio, and save the single-region matting map and feature summary in a TXT file. The visualization interface simultaneously displays the texture suppression map, enhancement difference map, damage mask, and final labeling map, allowing for intuitive comparison of detection results.
[0105] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for film structure damage detection and identification by fusing deep features with traditional texture analysis, characterized in that, include: Acquire images of the membrane structure to be detected; A first detection mode and a second detection mode are provided for users to choose from; In the first detection mode, the trained improved EAC-YOLO network is used to perform damage recognition on the image, outputting damage category, confidence level and bounding box coordinates, and the detection results are saved as a text file, which contains the category, confidence level and normalized bounding box coordinate information of each detected target; In the second detection mode, perform the following steps: Step 1: Based on the membrane texture type selected by the user, call the corresponding improved Gabor filter parameter group to suppress the background texture of the image; the texture type includes oblique texture, orthogonal texture, and sinusoidal curve texture; Step 2: Perform Local Structural Similarity (SSIM) calculation on the texture-suppressed image and the pre-stored non-destructive reference image to obtain the SSIM difference map. Then, perform gamma correction, contrast-limited adaptive histogram equalization (CLAHE), and Laplacian edge enhancement on the difference map in sequence. After weighted fusion, obtain the enhanced difference map. Step 3: Perform adaptive threshold segmentation on the enhanced difference map to obtain an initial damage mask; A texture energy difference map is calculated to generate an auxiliary mask. The initial damage mask is then fused with the auxiliary mask, and the final damage mask is obtained after morphological optimization. Step 4: Perform connected component analysis based on the final damage mask, extract the damage region, generate a damage matting map with a white background, and output the area, perimeter, circularity, and aspect ratio features of the damage region.
2. The method for membrane structure damage detection and identification based on the fusion of depth features and traditional texture analysis according to claim 1, characterized in that, The improved Gabor filter parameter set in step 1 includes: filters with directions of 0° and 90° and center frequencies of 0.1 to 0.2 for oblique textures; filters with directions of 45° and 135° and center frequencies of 0.1 to 0.2 for orthogonal textures; and filters with multiple directions and direction intervals of no more than 30° and center frequencies of 0.1 to 0.2 for sinusoidal textures. The Gaussian envelope scale of each parameter set is adjusted according to the period of the corresponding texture.
3. The method for membrane structure damage detection and identification based on the fusion of depth features and traditional texture analysis according to claim 1, characterized in that, The enhancement of the SSIM difference map in step 2 is specifically as follows: after normalizing the difference map, a gamma transform is performed to enhance the high-value region, and then CLAHE is used to improve the local contrast. At the same time, the edges of the difference map are extracted by the Laplacian operator. Finally, the CLAHE enhanced map and the Laplacian edge map are weighted and fused with a weight of 0.6:0.
4.
4. The method for membrane structure damage detection and identification based on the fusion of depth features and traditional texture analysis according to claim 1, characterized in that, The adaptive threshold segmentation in step 3 uses the Otsu method or the adaptive Gaussian thresholding method; the morphological optimization includes using elliptical structuring elements to perform closing and opening operations on the mask to remove noise and connect broken and damaged areas.
5. A method for membrane structure damage detection and identification that integrates depth features and traditional texture analysis according to any one of claims 1, characterized in that, In the second detection mode, the damage matting, damage marking map and the features are saved to a specified folder, and the intermediate processing results are displayed synchronously in the graphical user interface.
6. A system for film structure damage detection and identification based on the fusion of deep features and traditional texture analysis according to any one of claims 1-5, characterized in that, include: A first detection module, comprising: (1) EAC-YOLO network inference unit, used to load pre-trained weights or user-specified weight files, perform damage detection on input images, video streams or real-time camera footage, and output damage category, confidence level and location coordinates; (2) Result export unit, used to save the image detection results as a text file, wherein the text file contains the category, confidence level and normalized bounding box coordinates of each detected target; The second detection module includes: (1) Image preprocessing unit, used to perform grayscale conversion, Canny edge detection and binarization operations on the input image; (2) Texture suppression unit, which stores multiple sets of improved Gabor filter parameters corresponding to oblique texture, orthogonal texture and sinusoidal texture respectively, and is used to suppress background texture of the image according to the texture type selected by the user; (3) SSIM damage detection unit, used to perform local SSIM analysis on the texture-suppressed image and the reference image, and generate a damage mask through difference map enhancement, threshold segmentation and energy map fusion; (4) Damage extraction unit, used to extract the damage area according to the damage mask, generate a damage matting with a white background, and calculate the area, perimeter, roundness and aspect ratio of the damage area; (5) A graphical user interface, which is connected to the first detection module and the second detection module respectively, and provides detection mode switching, parameter setting, result display and result export functions.
7. The system for membrane structure damage detection and identification method integrating depth features and traditional texture analysis according to claim 6, characterized in that, The improved Gabor filter parameter group in the texture suppression unit is specifically: a filter group for oblique textures in the direction [0°, 90°]; a filter group for orthogonal textures in the direction [45°, 135°]; and a filter group for quasi-sine curve textures in the direction [0°, 30°, 60°, 90°, 120°, 150°]. The center frequency and scale of each filter group are preset according to the dominant frequency of the corresponding texture.
8. The system for membrane structure damage detection and identification method integrating depth features and traditional texture analysis according to claim 6, characterized in that, The SSIM damage detection unit generates a damage mask in the following manner: after calculating the SSIM difference map, it is sequentially subjected to gamma correction, CLAHE and Laplacian edge enhancement and weighted fusion to obtain an enhanced difference map; the enhanced difference map is then subjected to adaptive threshold segmentation to obtain an initial mask. Calculate the texture energy difference map between the image to be tested and the reference image and binarize it as an auxiliary mask; The initial mask is fused with the auxiliary mask, and then optimized by morphological closing and opening operations before being inverted to obtain the final mask with a white background and black damage.
9. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the method of any one of claims 1 to 5.
