Image data processing methods, apparatus, computer-readable storage media, and processors

Abstract

This invention discloses an image data processing method, apparatus, computer-readable storage medium, and processor. The method includes: acquiring image data of a subsurface diseased area; performing image enhancement and grayscale histogram equalization operations on the image data to obtain preprocessed image data; performing image recognition on the preprocessed image data to obtain a recognition result, wherein the recognition result includes semantic category labels corresponding to multiple pixels in the image data; determining diseased and non-diseased areas in the image data based on the semantic category labels corresponding to the multiple pixels in the recognition result; and performing three-dimensional visualization processing on the target location of the diseased area within the subsurface diseased area to obtain a detection result of the diseased area, wherein the detection result is used to represent the three-dimensional structural spatial features of the diseased area. This invention solves the technical problem of the difficulty in accurately detecting subsurface diseases.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method, apparatus, computer-readable storage medium, and processor for processing image data. Background Technology

[0002] Currently, ground-penetrating radar (GPR) is the mainstream method for detecting underground defects. However, its image segmentation techniques, such as threshold-based methods, region growing, edge detection, and graph cut-based segmentation, generally suffer from problems such as sensitivity to environmental conditions, insufficient ability to handle complex scenes, high computational resource requirements, and reliance on human experience. Furthermore, in 3D visualization, traditional techniques overly rely on experience in setting amplitude thresholds based on the subject matter, leading to subjectivity and uncertainty in the results. Therefore, there are technical challenges in accurately detecting underground defects.

[0003] There is currently no effective solution to the aforementioned technical problem of accurately detecting underground diseases. Summary of the Invention

[0004] This invention provides a method, apparatus, computer-readable storage medium, and processor for processing image data, to at least solve the technical problem of difficulty in accurately detecting underground diseases.

[0005] According to one aspect of the present invention, an image data processing method is provided. The method may include: acquiring image data of an underground diseased area; performing image enhancement and grayscale histogram equalization operations on the image data to obtain preprocessed image data; performing image recognition on the preprocessed image data to obtain a recognition result, wherein the recognition result includes semantic category labels corresponding to multiple pixels in the image data, the semantic category labels indicating the diseased area category to which the corresponding pixel belongs, the diseased area category including at least a diseased category and a non-diseased category; determining diseased areas and non-diseased areas in the image data based on the semantic category labels corresponding to multiple pixels in the recognition result; performing three-dimensional visualization processing on the target location of the diseased area within the underground diseased area to obtain a detection result of the diseased area, wherein the detection result represents the three-dimensional structural spatial features of the diseased area.

[0006] Optionally, image recognition is performed on the preprocessed image data to obtain recognition results, including: inputting the preprocessed image data into the target defect detection model for image recognition to obtain recognition results, wherein the target defect detection model is obtained by training an initial defect detection model with image data samples of the defect area.

[0007] Optionally, the target defect detection model includes an encoder and a decoder. The encoder consists of a multi-channel convolutional feature selection module, a parallel multi-directional attention mechanism, and a residual compression module. The multi-channel convolutional feature selection module is used to extract features at different scales in the image data. The parallel multi-directional attention mechanism is used to enhance the perception of features at different scales. The residual compression module is used to compress features at different scales. The decoder is used to reconstruct the image data based on the compressed features at different scales and generate recognition results based on the reconstructed image data.

[0008] Optionally, the image data processing method further includes: acquiring image data samples of the diseased area; labeling the image data samples to obtain label data corresponding to the image data samples; and training the initial defect detection model using the image data samples and the label data corresponding to the image data samples to obtain the target defect detection model.

[0009] Optionally, based on the semantic category labels corresponding to multiple pixels in the recognition result, the diseased area and non-diseased area in the image data are determined, including: determining the area corresponding to the pixel whose semantic category label is diseased in the recognition result as the diseased area; and determining the area corresponding to the pixel whose semantic category label is non-diseased in the recognition result as the non-diseased area.

[0010] Optionally, the target location of the diseased area within the underground diseased area is visualized in three dimensions to obtain the detection result of the diseased area. This includes: extracting the geometric information of the diseased area, wherein the geometric information includes at least the coordinates of the center point of the diseased area and the size information of the diseased area; determining the target location of the diseased area within the underground diseased area based on the coordinates of the center point of the diseased area and the size information of the diseased area; acquiring image data at the target location; reconstructing the amplitude information of the image data at the target location at different depths based on an amplitude reconstruction algorithm, wherein the amplitude reconstruction algorithm is constructed based on inverse divergence imaging technology and migration technology; and determining the detection result of the diseased area based on the amplitude information.

[0011] According to another aspect of the present invention, an image data processing apparatus is also provided. The apparatus may include: an acquisition unit for acquiring image data of an underground diseased area; a preprocessing unit for performing image enhancement and grayscale histogram equalization operations on the image data to obtain preprocessed image data; an identification unit for performing image recognition on the preprocessed image data to obtain an identification result, wherein the identification result includes semantic category labels corresponding to multiple pixels in the image data, the semantic category labels indicating the diseased area category to which the corresponding pixel belongs, and the diseased area category including at least a diseased category and a non-diseased category; a determination unit for determining diseased areas and non-diseased areas in the image data based on the semantic category labels corresponding to multiple pixels in the identification result; and a processing unit for performing three-dimensional visualization processing on the target location of the diseased area within the underground diseased area to obtain a detection result of the diseased area, wherein the detection result represents the three-dimensional structural spatial features of the diseased area.

[0012] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is run by a processor, it controls the device where the storage medium is located to execute the image data processing method of the present invention.

[0013] According to another aspect of the present invention, a processor is also provided. The processor is used to run a program, wherein the program executes the image data processing method of the embodiments of the present invention.

[0014] According to another aspect of the present invention, a computer program product is also provided. The program product includes computer instructions that, when executed by a processor, implement the image data processing method of the present invention.

[0015] In this embodiment of the invention, image data of underground diseased areas is identified to obtain identification results. These results include the identification of each pixel in the image data. Based on the identification results of each pixel, the diseased areas in the image data are determined. Then, the diseased areas are subjected to three-dimensional visualization processing to determine their three-dimensional spatial distribution characteristics. In other words, in this embodiment of the invention, pixel identification and three-dimensional visualization processing enhance the accuracy of diseased area identification, thereby solving the technical problem of poor detection accuracy for underground diseased areas. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0017] Figure 1This is a flowchart of an image data processing method according to an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of a Crack U-Net network structure according to an embodiment of the present invention;

[0019] Figure 3 This is a flowchart of an image recognition method based on the Crack U-Net model according to an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of a PDMA attention mechanism according to an embodiment of the present invention;

[0021] Figure 5 This is a structural diagram of an MC-FS module according to an embodiment of the present invention;

[0022] Figure 6 This is a structural diagram of a residual compression module according to an embodiment of the present invention;

[0023] Figure 7 This is a flowchart of a three-dimensional visualization imaging method according to an embodiment of the present invention;

[0024] Figure 8 This is a schematic diagram of an image data processing apparatus according to an embodiment of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, functional component, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, functional components, or devices.

[0027] According to an embodiment of the present invention, an embodiment of an image data processing method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0028] Figure 1 This is a flowchart of an image data processing method according to an embodiment of the present invention, such as... Figure 1 As shown, the method may include the following steps:

[0029] Step S101: Collect image data of the underground diseased area.

[0030] In the technical solution provided in step S101 of the present invention, high dynamic ground penetrating radar (GPR) equipment is typically used to acquire image data of underground disease areas. GPR is a non-destructive detection technology that can emit high-frequency electromagnetic waves and receive reflected signals. By analyzing these signals, two-dimensional or three-dimensional images of underground structures are generated, thereby detecting underground diseases.

[0031] In this embodiment, to more comprehensively cover the underground diseased area and acquire high-resolution data, the survey lines of the GPR ground-penetrating radar need to be arranged in a grid pattern, such as a "well" shape. This arrangement ensures that image data is obtained from different directions, thereby improving the accuracy and reliability of disease detection.

[0032] Optionally, a GPR device is used to collect data along a preset survey line. The device continuously emits electromagnetic waves and receives reflected signals to collect image data of the diseased area. After processing, these signals are converted into a two-dimensional image (B-scan), where each row represents profile information along the survey line direction, and each column represents information at different depths.

[0033] Step S102: Perform image enhancement and grayscale histogram equalization operations on the image data to obtain preprocessed image data.

[0034] In the technical solution provided by step S102 of the present invention, image enhancement is a process of improving image quality by adjusting attributes such as brightness, contrast, and color. In the processing of underground disease images, the disease features may not be obvious enough in the original image, or the image quality may be poor due to limitations in the signal strength of the ground-penetrating radar equipment and environmental factors (such as soil type and humidity). Therefore, by adjusting the image contrast and brightness to enhance the difference between the diseased area and the surrounding normal structure, it is helpful to more accurately identify disease features.

[0035] Optionally, grayscale histogram equalization is an image processing technique used to improve image contrast, especially when the grayscale distribution of an image is uneven. In ground-penetrating radar images, grayscale histogram equalization can enhance details in dark areas of the image, making information from different grayscale levels more balanced, thereby improving the visibility of disease features.

[0036] Optionally, by performing image enhancement and grayscale histogram equalization on the image data, the quality of the defect image can be improved, making its features more prominent and its contrast higher. This reduces the processing difficulty of the target defect detection model (Crack U-Net) and improves the model's segmentation accuracy and the accuracy of 3D visualization processing. The Crack U-Net model is obtained by improving the U-Net model.

[0037] Step S103: Perform image recognition on the preprocessed image data to obtain the recognition result.

[0038] In the technical solution provided in step S103 of the present invention, the preprocessed image data is image recognized by a deep learning model target defect detection model, and the recognition result includes the semantic category label of each pixel in the image.

[0039] In this embodiment, preprocessed image data is input into the target defect detection model. The image data enters the encoder part of Crack U-Net, where the encoder extracts features from the image using a multi-layer convolutional neural network (CNN). The encoder includes a multi-convolution-features selection module (MC-FS), a parallel multi-directional attention mechanism (PMDA), and a residual squeeze block (RS) module. These modules work together to extract defect features from different scales and angles, enhancing the model's ability to identify defect details.

[0040] Optionally, features at different scales can be extracted through parallel standard convolution and dilated convolution, while a channel attention mechanism is used to weight the extracted features to ensure that the extracted features are highly correlated with the disease.

[0041] Optionally, a parallel multi-directional attention mechanism is used to establish relationships between pixels, enhancing the network's ability to extract crack features. In particular, when dealing with long and continuous cracks, it can more effectively identify and highlight crack pixels, thereby improving segmentation accuracy.

[0042] Optionally, the residual compression module compresses the shallow features of the encoder before feature connection to remove "coarse" information, reduce the number of network parameters, and ensure that the feature information received by the decoder is as effective as possible.

[0043] Optionally, the features extracted by the encoder are passed to the decoder, which uses upsampling and feature fusion techniques to progressively restore the image resolution, combining the high-level features of the encoder with the low-level features to ultimately generate a recognition result with the same size as the input image. This recognition result contains a semantic category label for each pixel, indicating whether each pixel belongs to a diseased area and what type of disease it belongs to (e.g., cracks, cavities, delamination, etc.).

[0044] Optionally, the semantic category labels output by the model can be post-processed, such as through denoising and boundary refinement, to obtain a clear segmentation map of the diseased area. By comparing the segmentation map with the original image, the location and extent of the diseased area can be intuitively identified, providing accurate two-dimensional location information for subsequent 3D visualization and disease analysis.

[0045] In this step, throughout the image recognition process, the Crack U-Net model utilizes deep learning techniques, especially an improved attention mechanism and feature selection module, to automatically learn and extract disease features from the image, thereby achieving high-precision image segmentation. The resulting recognition results can not only distinguish between diseased and non-diseased areas but also provide classification information on disease types, which is of great significance for the health assessment and maintenance strategy formulation of underground structures.

[0046] Step S104: Based on the semantic category labels corresponding to multiple pixels in the recognition results, determine the diseased and non-diseased areas in the image data.

[0047] In the technical solution provided by step S104 of the present invention, as can be seen from the description of step S102 above, the recognition result includes semantic category labels corresponding to multiple pixels in the image data. Based on this, the diseased area and non-diseased area in the image data can be determined according to the semantic category labels corresponding to multiple pixels in the recognition result.

[0048] Optionally, based on the recognition results of the Crack U-Net model, each pixel corresponds to a label of disease category or non-disease category. By analyzing these labels, pixels belonging to diseased areas and pixels belonging to non-disease areas can be determined.

[0049] Optionally, the region corresponding to the pixel labeled as a disease category in the identification result is determined as the disease region, and the region corresponding to the pixel labeled as a non-disease category in the identification result is determined as the non-disease region.

[0050] For example, pixels with specific category labels are clustered to form continuous disease area boundaries. For instance, if the model predicts that a pixel belongs to the crack category, then that pixel and its neighboring pixels with the same label will be marked as part of the crack disease area.

[0051] In this step, based on the semantic category label of each pixel, the diseased and non-diseased areas in the image data can be accurately identified, providing a precise data foundation for subsequent 3D visualization processing and disease assessment.

[0052] Step S105: Perform three-dimensional visualization processing on the target location of the diseased area within the underground diseased area to obtain the detection results of the diseased area.

[0053] In the technical solution provided by step S105 of the present invention, the target location of the diseased area in the underground diseased area is subjected to three-dimensional visualization processing to obtain the detection result of the diseased area. That is, the semantic category label output by the improved image segmentation algorithm is used, combined with the two-dimensional image data collected by the high dynamic ground penetrating radar, and the accurate three-dimensional model of the diseased area is generated through the true amplitude reconstruction algorithm and three-dimensional visualization technology.

[0054] In this embodiment, after segmenting the underground disease image and determining the disease area, a true amplitude reconstruction algorithm can be applied to enhance the visualization of disease features. Based on inverse scattering imaging and migration techniques, the true amplitude reconstruction algorithm corrects and recovers the amplitude energy of the wave field at the scattering source location, thus more accurately representing the true state of the disease area in the two-dimensional image, improving image quality and contrast, and making the disease features more prominent.

[0055] Optionally, by fitting and interpolating B-scan data at different time depths, lateral positions, and longitudinal positions, the two-dimensional image can be converted into a three-dimensional space. This process, by analyzing and integrating B-scan data from multiple directions, can construct a complete three-dimensional spatial structure of the diseased area, including its accurate depth, lateral and longitudinal distribution, as well as the shape and size of the disease.

[0056] Optionally, after the 3D visualization process is completed, 3D slice images of the affected area can be obtained. These images provide views of the disease from different angles and depths, which helps to more comprehensively understand and analyze the nature and impact of the disease. Through 3D slices, the burial depth of the anomaly and its distribution in 3D space can be accurately inferred, which is crucial for the design of subsequent remediation plans and the formulation of monitoring plans.

[0057] Optionally, the detection results, i.e., the detection information of the diseased area after 3D visualization processing, are used to represent the 3D structure and spatial characteristics of the diseased area. This information includes, but is not limited to, the precise geometry of the diseased area, its precise location in underground space, its depth, and its possible direction of expansion. The 3D visualization results not only provide an intuitive image of the disease but also represent the parameters of the disease in data form, facilitating more in-depth analysis and decision-making by professionals.

[0058] In steps S101 to S105 above, the image data of the underground diseased area is identified to obtain an identification result. This identification result includes the identification result of each pixel in the image data. Based on the identification result of each pixel, the diseased area in the image data is determined. Then, the diseased area is subjected to three-dimensional visualization processing to determine the three-dimensional spatial distribution characteristics of the diseased area. In other words, in this embodiment of the invention, pixel identification and three-dimensional visualization processing can enhance the identification accuracy of the diseased area, thereby solving the technical problem of poor detection accuracy of underground diseased areas.

[0059] The method described in this embodiment will be further described below.

[0060] As an optional embodiment, step S103, performing image recognition on the preprocessed image data to obtain recognition results, includes: inputting the preprocessed image data into the target defect detection model for image recognition to obtain recognition results, wherein the target defect detection model is obtained by training an initial defect detection model with image data samples of the defect area.

[0061] In this embodiment, the target defect detection model can be the Crack U-Net model, which is an improved and optimized version of the U-Net architecture based on deep learning.

[0062] Optionally, the Crack U-Net defect detection model, based on U-Net, introduces a parallel multi-directional attention mechanism (PMDA), a multi-path convolution-feature selection module (MC-FS), and a residual compression module (RS). PMDA enhances the model's extraction of crack features by establishing relationships between pixels, especially when dealing with long and continuous cracks, enabling more accurate capture and enhancement of crack pixels. The MC-FS module increases feature diversity through parallel convolution and channel attention mechanisms, thereby improving feature effectiveness and ensuring the model extracts only key features relevant to the defect. The RS module compresses shallow features of the encoder before feature concatenation, removing "coarse" information, reducing the number of model parameters, avoiding noise introduction into deep features, and improving operational efficiency and feature effectiveness.

[0063] Optionally, the target defect detection model (Crack U-Net) is trained based on a large number of labeled image samples of diseased areas. These samples include various typical forms and scenarios of underground defects, such as different types of cracks, cavities, and voids. The model optimizes its parameters by minimizing the difference between the predicted results and the true labels (typically using the cross-entropy loss function). During training, the model gradually learns how to effectively distinguish between diseased and non-diseased areas, and how to identify different types of defects. After training, the model can automatically identify and segment diseased areas based on the input image data, and output a semantic category label for each pixel.

[0064] Optionally, after training the target defect detection (Crack U-Net) model, the preprocessed image can be input into the trained Crack U-Net model. The encoder part of the model extracts defect features from the image through MC-FS and PMDA mechanisms, and weights crack pixels and other defect areas to enhance the model's ability to perceive and capture these features. The features extracted by the encoder are fused and upsampled in the decoder to gradually restore the image resolution and finally generate a recognition result consistent with the input image size. In this process, the RS module compresses the features to ensure that the feature information received by the decoder is the most effective, avoiding unnecessary computation and information redundancy.

[0065] Optionally, the recognition results output by the Crack U-Net model include semantic category labels for each pixel. These labels indicate whether each pixel belongs to a diseased area and, if so, which disease category it belongs to. The accuracy of the recognition results depends on the quality of model training and the clarity of the preprocessed image. Through the deep learning capabilities of the Crack U-Net model, high-precision segmentation and classification of underground diseases can be achieved, providing accurate two-dimensional diseased area information for subsequent three-dimensional visualization processing.

[0066] In summary, image recognition of the preprocessed image data is achieved through the application of the Crack U-Net model. This model, using deep learning technology and combining innovative PMDA, MC-FS, and RS modules, automatically learns and extracts disease features from images, outputting high-precision disease area identification results. This process not only improves the accuracy of image segmentation but also reduces the reliance on manually set thresholds or seed points, greatly enhancing work efficiency and the reliability of results.

[0067] As an optional implementation, the target defect detection model includes an encoder and a decoder. The encoder consists of a multi-channel convolutional feature selection module, a parallel multi-directional attention mechanism, and a residual compression module. The multi-channel convolutional feature selection module is used to extract features at different scales in the image data. The parallel multi-directional attention mechanism is used to enhance the perception of features at different scales. The residual compression module is used to compress features at different scales. The decoder is used to reconstruct the image data based on the compressed features at different scales and generate recognition results based on the reconstructed image data.

[0068] In this embodiment, the target defect detection model (Crack U-Net) achieves high-precision segmentation of underground disease images through multi-path convolutional feature selection, parallel multi-directional attention mechanism, and residual compression in the encoder, as well as feature fusion and image reconstruction in the decoder. The model not only accurately identifies disease features but also maintains high efficiency during processing, reducing unnecessary computational resource consumption and providing accurate image recognition results for subsequent 3D visualization and other analyses.

[0069] As an optional implementation, the image processing method further includes: acquiring image data samples of the diseased area; labeling the image data samples to obtain label data corresponding to the image data samples; and training an initial defect detection model using the image data samples and the label data corresponding to the image data samples to obtain a target defect detection model.

[0070] In this embodiment, image data samples of the diseased area are acquired and labeled. Then, the labeled data is used to train the initial defect detection model, ensuring that the model can learn the features of the diseased area and perform accurate segmentation.

[0071] Optionally, a large number of underground disease image data samples can be collected. These samples can be acquired at different underground disease detection sites using high-dynamic ground-penetrating radar equipment, ensuring that the samples cover various types of diseases, such as cracks, cavities, and voids, as well as disease images under different conditions, such as images at different depths, geological backgrounds, and under different lighting and noise levels. Such a sample set enables more comprehensive model training and improves its generalization ability in practical applications.

[0072] Optionally, after acquiring the disease image data samples, each collected image data sample is labeled, that is, specifying which pixels belong to the diseased area and which belong to the non-diseased area. The labeling result will assign a semantic category label to each pixel, such as crack category, cavity category, or non-disease category. This process is the foundation of model training, ensuring that the model can learn the correct disease features.

[0073] Optionally, the Crack U-Net model needs to be pre-trained before starting the training process, for example, using model parameters that have performed well on other related tasks as initialization. Pre-training can accelerate model convergence and improve training efficiency. Once the initial model is ready, it can be trained using a labeled dataset.

[0074] Optionally, the Crack U-Net model is trained using labeled image data samples and corresponding label data. The training process includes two stages: forward propagation and back propagation. In forward propagation, the input image data passes through the model, outputting the corresponding predicted label; in back propagation, the difference (loss) between the predicted label and the actual label is calculated, and the model parameters are adjusted based on this loss to reduce prediction error. The goal of training is to optimize the model parameters so that the model achieves the minimum loss on the test data, that is, the model's prediction results are as close as possible to the manually labeled real results.

[0075] Optionally, during training, the model's hyperparameters, such as learning rate, batch size, and optimizer selection, need to be tuned to optimize the training process and model performance. Furthermore, the model should be evaluated using an independent validation dataset to ensure that it not only performs well on the training data but also has the ability to maintain stable performance on unseen data.

[0076] Optionally, after training, the model needs to undergo a final performance evaluation on a test dataset. The test dataset should contain disease images different from those in the training dataset to verify the model's generalization ability. If the test results are unsatisfactory, it may be necessary to return to the training phase, adjust the model structure or parameters, and repeat the training process until the model achieves satisfactory performance on the test data.

[0077] Through the above steps, a fully trained and optimized Crack U-Net model, i.e., a target defect detection model, can be obtained. This model can automatically identify and segment defect areas in underground defect images without manual intervention. The generated target model is the core of the entire technical solution, used for subsequent defect image segmentation and 3D visualization processing.

[0078] As an optional implementation method, based on the semantic category labels corresponding to multiple pixels in the recognition result, the diseased area and non-diseased area in the image data are determined, including: determining the area corresponding to the pixel with the semantic category label of disease in the recognition result as the diseased area; and determining the area corresponding to the pixel with the semantic category label of non-disease in the recognition result as the non-diseased area.

[0079] In this embodiment, the identification results of diseased and non-diseased areas in the image data can be determined based on the semantic category label corresponding to each pixel in the identification results.

[0080] Optionally, in the recognition results, the regions corresponding to pixels labeled as disease categories by the model (i.e., whose semantic category label is disease category) will be determined as disease regions. The regions corresponding to pixels labeled as non-disease categories by the model (i.e., whose semantic category label is non-disease category) will be determined as non-disease regions.

[0081] Optionally, after initially identifying the affected and non-affected areas, post-processing steps are usually required to optimize the identification results. For example, morphological operations (such as dilation and erosion) can be used to smooth the boundaries of affected areas and eliminate isolated pixels that may be caused by noise. Furthermore, connected component analysis can be used to identify and label complete affected areas, ensuring the consistency and completeness of the detection results.

[0082] In this step, by identifying the semantic category label of each pixel included in the result, the diseased area and non-diseased area in the image can be accurately distinguished, thus improving the accuracy of disease detection.

[0083] As an optional embodiment, step S105 involves performing three-dimensional visualization processing on the target location of the diseased area within the underground diseased area to obtain the detection result of the diseased area. This includes: extracting the geometric information of the diseased area, wherein the geometric information includes at least the coordinates of the center point of the diseased area and the size information of the diseased area; determining the target location of the diseased area within the underground diseased area based on the coordinates of the center point of the diseased area and the size information of the diseased area; acquiring image data at the target location; reconstructing the amplitude information of the image data at the target location at different depths based on an amplitude reconstruction algorithm, wherein the amplitude reconstruction algorithm is constructed based on inverse divergence imaging technology and offset technology; and determining the detection result of the diseased area based on the amplitude information.

[0084] In this embodiment, performing three-dimensional visualization processing on the diseased area to obtain detection results is a complex process based on multiple steps. It involves extracting geometric information of the diseased area from two-dimensional image data, determining the precise location of the disease in the underground space, and using an amplitude reconstruction algorithm to generate amplitude information of the diseased area at different depths, thereby ultimately determining the three-dimensional structure and spatial characteristics of the diseased area.

[0085] Optionally, for 3D visualization, it is necessary to extract the geometric information of the diseased area from the recognition results, including but not limited to the center point coordinates and size information of the diseased area. The center point coordinates are obtained by calculating the average position of the pixels in the diseased area, while the size information is determined by analyzing the boundary of the diseased area, including the length and width of the disease.

[0086] Optionally, after determining the geometric information of the affected area, the target location of the disease within the underground affected area can be determined by combining the two-dimensional image data acquired by ground-penetrating radar. This typically involves converting the two-dimensional coordinates into a relative position in the underground three-dimensional space, using the coordinates and size information of the center point of the affected area, combined with the propagation path and reflection characteristics of the radar signal, to calculate the depth and specific location of the disease in the underground space.

[0087] Optionally, after determining the target location of the damage in the underground space, these locations can be surveyed in more detail. This may involve redeploying survey lines, using a denser survey grid, or adjusting the parameters of radar equipment to obtain higher-quality image data. This image data will be used for subsequent amplitude information reconstruction and 3D visualization.

[0088] Optionally, to more accurately represent the characteristics of the disease, it is necessary to reconstruct the amplitude information of the acquired image data of the target location. This process is based on inverse scattering imaging and migration techniques. By recovering the true amplitude energy of the wave field at the scattering source location, the amplitude distribution of the diseased area at different depths can be obtained. Inverse scattering imaging can recover the true shape of the object from the scattered waves, while migration techniques can correct for the positional shift of the scattered waves caused by factors such as changes in wave velocity or topographic undulations, ensuring the accuracy and consistency of the amplitude information.

[0089] Optionally, based on the reconstructed amplitude information, three-dimensional visualization processing can be performed. By fitting B-scan data at different time depths, lateral positions, and longitudinal positions, accurate interpolation of limited B-scan data is achieved, enabling precise three-dimensional slice imaging of the diseased area. In this process, the 3D module of the ReflexW software is used to process the two-dimensional image data, generating three-dimensional images of the diseased area at different depths and directions, thereby constructing a complete three-dimensional model of the diseased area, including its accurate burial depth, lateral and longitudinal distribution, as well as the shape and size of the disease.

[0090] Optionally, the results of the 3D visualization processing, i.e., the detection results of the diseased area, are used to represent the 3D structure and spatial characteristics of the diseased area. This information not only provides an intuitive image of the disease, but also represents the parameters of the disease in data form, such as burial depth, volume, and location coordinates, which facilitates more in-depth analysis and decision-making by professionals.

[0091] In this step, the three-dimensional visualization of the target location of the diseased area within the underground diseased area is a complex process involving multiple steps, including data acquisition, feature extraction, geometric information analysis, amplitude information reconstruction, and three-dimensional image generation. Through this series of steps, an accurate three-dimensional model of the diseased area can be obtained, providing a numerical reference for subsequent disease assessment. This not only improves the accuracy of disease detection but also reduces reliance on expensive equipment, enhancing the efficiency and economy of the entire detection process.

[0092] The technical solutions of the embodiments of the present invention will be illustrated below with reference to preferred embodiments.

[0093] Currently, existing methods for segmenting underground disease images mainly include threshold-based segmentation, region growing, edge detection, graph cut-based segmentation, and deep learning-based segmentation. Traditional 3D visualization methods primarily set amplitude thresholds for the synthesized 3D data volume, displaying only high-amplitude reflection signals to extract the geometric structure and spatial distribution features of the underground target. However, threshold-based segmentation is sensitive to illumination changes and noise; when the image is affected by illumination changes or noise interference, the segmentation effect may be compromised, leading to inaccurate segmentation. For complex scenes where the grayscale difference between the target and the background is not significant or where multiple grayscale levels exist, threshold segmentation may not effectively distinguish different regions. Region growing relies on the selection of seed points and the setting of similarity criteria, and may not accurately segment discontinuous or feature-similar regions. While edge detection can provide shape and contour information of the target, it is easily affected by noise and image blurring, leading to inaccurate edge detection or the generation of false edges. Graph cut-based segmentation, although capable of handling complex scenes, has a high computational cost and may not be suitable for real-time processing or large-scale image data.

[0094] Furthermore, traditional 3D visualization imaging relies on the operator's experience in setting amplitude thresholds, leading to inconsistent results depending on the threshold setting. Additionally, ground-penetrating radar (GPR) systems with built-in visualization capabilities are expensive, often several times the price of ordinary GPR, and their visualization quality is generally mediocre, again relying on the interpreter's experience. This presents a technical challenge in accurately detecting underground defects.

[0095] However, this invention proposes a three-dimensional processing method for segmentation and visualization of underground disease images. By introducing a more advanced image segmentation algorithm, it improves the accuracy and robustness of segmentation, enabling stable segmentation results under different lighting conditions and noise environments. Utilizing the multi-layered, non-linear feature extraction capabilities of deep learning models, it aims to better handle the complex shapes and textures of underground diseases, achieving more accurate segmentation. Through automated algorithms and intelligent processing technologies, it reduces manual intervention, increases the automation level of the entire process, and thus improves work efficiency and accuracy.

[0096] The following section will further introduce the underground disease images and three-dimensional visualization imaging technology in the embodiments of the present invention.

[0097] In this embodiment, the image segmentation algorithm is an improvement on the U-Net network structure, named Crack U-Net. Figure 2 This is a schematic diagram of a Crack U-Net network structure according to an embodiment of the present invention. Figure 2 As shown, the Crack U-Net network is divided into two parts: the left half is the encoder, and the right half is the decoder. The encoder part is further divided into five parts, each containing a multi-path convolutional feature selection (MC-FS) module, a parallel multi-directional attention mechanism (PMDA) module, and a residual compression (RS) module. The input image size is 512×512 to facilitate downsampling by a factor of 2 and to keep the size as close as possible to the original image size, preventing feature loss due to image reduction. The MC-FS module is placed first in each module to increase the number of channels while extracting multiple features. These features are then selected to ensure the effectiveness of the input features in subsequent operations, thereby improving the encoder's encoding ability. The model chooses to use 3×3 convolutional kernels for further feature extraction because the MC-FS module includes dilated convolutions, which already have a large receptive field. Using smaller convolutional kernels reduces the overall number of parameters in the model. Next, the parallel multi-directional attention mechanism module establishes relationships between feature pixels, thereby weighting the crack pixels and strengthening the network's ability to extract crack features. Finally, max pooling is performed on the output features to reduce their resolution. This process is repeated five times to complete feature encoding. Starting from the last layer of the encoder, the features enter the decoder. The output features of the last encoder layer are upsampled, doubling the resolution to form the features in the decoder. These features are then concatenated with the corresponding encoder output features. At this point, the encoder features first undergo channel compression in the compression module to remove unnecessary shallow features and reduce the network size before concatenation, preventing information loss. Then, feature extraction is performed twice. This process is repeated four times to complete decoding. The resolution of the decoded feature map is the same as the original. Figure 1The number of channels is equal to the number of categories. Finally, all channels of each pixel are activated to complete pixel-level classification.

[0098] The following section will further explain the process of image recognition using the Crack U-Net model.

[0099] Figure 3 This is a flowchart of an image recognition method based on the Crack U-Net model according to an embodiment of the present invention, such as... Figure 3 As shown, the method includes the following steps:

[0100] Step S301, data preprocessing.

[0101] In this embodiment, the input underground disease image is preprocessed, including image enhancement and grayscale histogram equalization, to improve the image quality and contrast, providing a better foundation for subsequent segmentation.

[0102] Step S302: Construct the Crack U-Net model.

[0103] In this embodiment, an improved U-Net model, namely Crack U-Net, is constructed. This model, based on U-Net, introduces a parallel multi-directional attention mechanism (PMDA), a multi-path convolution-feature selection module (MC-FS), and a residual compression (RS) module to enhance the network's ability to perceive features at different scales and improve its segmentation performance.

[0104] Optionally, cracks typically exhibit the following morphological characteristics: cracks are generally long, thin, and continuous. Therefore, in radar images, when a pixel is a crack pixel, its neighboring pixels are highly likely to also contain crack pixels; if this pixel is background, its adjacent areas are also highly likely to be background. Therefore, establishing a correlation between pixels based on this characteristic is more beneficial for crack feature extraction. However, since channel attention mechanisms weight features along the channel dimension, they are not suitable for extracting crack features with correlations between pixels, and current spatial attention mechanisms do not include a pixel correlation mechanism. Therefore, a parallel multi-directional attention mechanism (PMDA) module is proposed to establish a pixel correlation mechanism to improve the network's ability to extract crack features.

[0105] Optionally, PMDA is a spatial attention mechanism used in the Crack U-Net encoder. Its principle is to decompose a 3×3 convolution kernel into 4 parts, each part being a pixel that establishes a relationship with its neighboring pixels. Figure 4 This is a schematic diagram of a PDMA attention mechanism according to an embodiment of the present invention, such as... Figure 4As shown, this module is divided into four paths. Looking from top to bottom, for an input feature A of shape H×W×C, the first two paths use a 1×2 and a 2×1 convolutional kernel respectively to establish a relationship between each pixel in feature A and one pixel in its horizontal and vertical directions. These two convolutions yield features B and C respectively. The third and fourth paths use a 2×2 convolutional kernel with an expansion ratio of 2 and a 1×1 convolutional kernel. These two paths work together to establish a relationship between each pixel and its four adjacent pixels on its two diagonals. Therefore, the results of these two paths are added together to obtain the fused feature D. Then, the concatenation result F of features B, C, and D is combined using a 2×2 convolutional kernel to merge all channels of feature F to obtain a single-channel feature map G. An activation function is then used to convert its values ​​into weights in the (0, 1) interval. Finally, G is multiplied by the original feature map A to complete the pixel weighting in the feature map, resulting in A′. Since only 5 convolutional kernels are used and the kernels are very small, the parallel multi-directional attention module is also a lightweight network module, requiring only a few additional network parameters and is easy to implement.

[0106] Optionally, due to complex road conditions and the noise generated by ground-penetrating radar itself, radar images may contain signal interference from other conditions and equipment noise. This can cause the network to incorrectly extract features from noise or other conditions, reducing the effectiveness of the extracted features and consequently reducing the network's segmentation ability. To address this, a Multi-Path Convolutional Feature Selection (MC-FS) module is proposed to improve the effectiveness of the features extracted by the network. The principle of the MC-FS module is to ensure effectiveness through feature diversity; that is, multiple features are first extracted, and then a channel attention mechanism is used to weight the channels to select the more effective features.

[0107] Optionally, the MC-FS module consists of parallel standard convolutions, dilated convolutions, and channel attention modules, and is used in the encoder of Crack U-Net, just like the PMDA module. Figure 5 This is a structural diagram of an MC-FS module according to an embodiment of the present invention, as shown below. Figure 5As shown in the figure, w, h, and c represent the width, height, and number of channels of the input feature, respectively, and C represents the number of output channels. The input feature is first subjected to four parallel convolutions, two of which are 3×3 and 2×2 dilated convolutions with an expansion rate of 2. This aims to obtain a larger receptive field with fewer parameters. The addition of the standard 3×3 and 2×2 convolutions avoids the checkerboard effect of dilated convolutions, eliminating the discontinuity and local feature loss problems in feature extraction. Therefore, for one input feature, the MC-FS module generates four features, and the number of channels for each feature is 1 / 4 of the output feature. Connecting these four features restores the number of output channels. Finally, a channel attention module is added to weight each channel, achieving feature selection and assigning higher weights to effective channels.

[0108] Optionally, the use of the above modules inevitably increases the number of network parameters, which to some extent affects the network's operating efficiency. Furthermore, traditional U-Net suffers from the following unreasonable feature connection design: the decoder connects to features from the corresponding encoder layer as supplementary information. However, the encoder, compared to the decoder, belongs to a shallower region and contains a large amount of "coarse" information. Connecting all these features not only significantly increases the number of network parameters but also forces the decoder to re-"refine" these features containing unnecessary or "coarse" information when supplementing information, which is tantamount to adding noise to deeper features. Therefore, to reduce the number of network parameters and increase the effectiveness of the feature connection from the encoder to the decoder in Crack U-Net, a residual compression module (RS) is proposed. The RS module reduces the number of network model parameters by removing redundant feature channels and is used before the shallow features of the encoder are residually connected to the decoder.

[0109] Alternatively, in Crack U-Net, encoder features are supplemented with appropriate feature information for the decoder after passing through the RS module, without having to connect all features into the decoder. Figure 6 This is a structural diagram of a residual compression module according to an embodiment of the present invention. After the features enter the compression module, they undergo max pooling with a stride of 2 to retain the feature value with the largest response. Then, unpooling is used to restore the resolution. The purpose of this is to retain the spatially most efficient features for encoder connections. Finally, a 2×2 ordinary convolution is used to change the feature channels. Here, the compression rate is 0.25, which means the number of channels is compressed to 1 / 4 of the original. Therefore, the RS module is also a lightweight network module.

[0110] Step S303, Model Training.

[0111] In this embodiment, the Crack U-Net model is trained using a large amount of labeled underground disease image data. The model parameters are continuously adjusted by optimizing the loss function until a satisfactory segmentation effect is achieved.

[0112] Step S304, image segmentation.

[0113] In this embodiment, the preprocessed image of underground disease is input into the trained Crack U-Net model for image segmentation. The model outputs a semantic category label for each pixel, thereby achieving accurate segmentation of the diseased area.

[0114] Step S305, post-processing.

[0115] In this embodiment, the identification results are post-processed, such as extracting the bounding rectangle of the diseased area, recording the coordinates of the center point, length and width, etc., to facilitate subsequent analysis and processing.

[0116] In steps S301 to S305 above, by training the Crack U-Net model and then using the trained Crack U-Net model to segment the image, the semantic category label of each pixel in the image data can be obtained, thereby achieving accurate segmentation of the diseased area in the image data.

[0117] Optionally, after obtaining the accurate segmentation results of the diseased area in the image data, the segmentation results can be visualized in three dimensions.

[0118] The following section will further introduce the 3D visualization imaging method.

[0119] Figure 7 This is a flowchart of a three-dimensional visualization imaging method according to an embodiment of the present invention, such as... Figure 7 As shown, the method includes the following steps:

[0120] Step S701, data acquisition.

[0121] In this embodiment, the abnormal area is re-detected, and the survey lines are arranged in a grid pattern (generally in a "well" shape); then, a high dynamic ground-penetrating radar is used to collect data and obtain two-dimensional radar image data.

[0122] Step S702, data processing.

[0123] In this embodiment, the 2D module of the application software performs simple static correction and gain processing on all survey line data.

[0124] Step S703, true amplitude reconstruction.

[0125] In this embodiment, based on the single-channel signal of the anomalous body, a reconstruction algorithm for the true amplitude is established using the inverse scattering imaging method and the migration method to recover the amplitude energy of the wave field at the scattering source location. Then, based on the B-scan data, the interface of the anomalous body is restored to its true state from point to surface.

[0126] Step S704, 3D visualization.

[0127] In this embodiment, B-scan data at different time depths, lateral positions, and longitudinal positions are fitted to complete the true interpolation of limited B-scan data, thereby realizing true three-dimensional slice imaging of the anomaly and more accurately inferring the burial depth and distribution of the anomaly.

[0128] Step S705, result verification.

[0129] In this embodiment, the 3D visualization results are compared and analyzed with the drilling verification results to verify the accuracy and reliability of the 3D visualization technology.

[0130] In steps S701 to S705 above, the intuitiveness of the results and the scientific nature of the decision-making are enhanced by three-dimensional visualization. At the same time, the reliability and application value of the entire detection process are improved through rigorous result verification.

[0131] According to embodiments of the present invention, an image data processing apparatus is also provided. It should be noted that this image data processing apparatus can be used to execute the image data processing method described in the embodiments.

[0132] Figure 8 This is a schematic diagram of an image data processing apparatus according to an embodiment of the present invention. Figure 8 As shown, the network configuration strategy adjustment device 800 may include: a collection unit 801, a preprocessing unit 802, an identification unit 803, a determination unit 804, and a processing unit 805.

[0133] Acquisition unit 801 is used to acquire image data of underground diseased areas;

[0134] The preprocessing unit 802 is used to perform image enhancement and grayscale histogram equalization operations on the image data to obtain preprocessed image data.

[0135] The recognition unit 803 is used to perform image recognition on the preprocessed image data to obtain recognition results. The recognition results include semantic category labels corresponding to multiple pixels in the image data. The semantic category labels are used to indicate the disease area category to which the corresponding pixel belongs. The disease area category includes at least disease category and non-disease category.

[0136] The determining unit 804 is used to determine the diseased area and non-diseased area in the image data based on the semantic category labels corresponding to multiple pixels in the recognition result;

[0137] The processing unit 805 is used to perform three-dimensional visualization processing on the target location of the diseased area in the underground diseased area to obtain the detection result of the diseased area, wherein the detection result is used to represent the three-dimensional structural spatial characteristics of the diseased area.

[0138] Optionally, the recognition unit 803 is further configured to: input the preprocessed image data into the target defect detection model for image recognition and obtain the recognition result, wherein the target defect detection model is obtained by training the initial defect detection model with image data samples of the defect area.

[0139] Optionally, the image data processing device 800 is further configured to: acquire image data samples of the diseased area; annotate the image data samples to obtain label data corresponding to the image data samples; and train the initial defect detection model using the image data samples and the label data corresponding to the image data samples to obtain the target defect detection model.

[0140] Optionally, the determining unit 804 is further configured to: determine the region corresponding to the pixel whose semantic category label is disease category in the recognition result as disease region; and determine the region corresponding to the pixel whose semantic category label is non-disease category in the recognition result as non-disease region.

[0141] Optionally, the processing unit 805 is further configured to: extract geometric information of the diseased area, wherein the geometric information includes at least the coordinates of the center point of the diseased area and the size information of the diseased area; determine the target location of the diseased area in the underground diseased area based on the coordinates of the center point of the diseased area and the size information of the diseased area; acquire image data at the target location; reconstruct the amplitude information of the image data at the target location at different depths based on an amplitude reconstruction algorithm, wherein the amplitude reconstruction algorithm is constructed based on inverse divergence imaging technology and migration technology; and determine the detection result of the diseased area based on the amplitude information.

[0142] In this embodiment, image data of underground diseased areas is identified to obtain identification results. These results include the identification of each pixel in the image data. Based on the identification results of each pixel, the diseased areas in the image data are determined. Then, the diseased areas are subjected to three-dimensional visualization processing to determine their three-dimensional spatial distribution characteristics. In other words, in this embodiment of the invention, pixel identification and three-dimensional visualization processing can enhance the accuracy of diseased area identification, thereby solving the technical problem of poor detection accuracy of underground diseased areas.

[0143] According to an embodiment of the present invention, a computer-readable storage medium is also provided, the storage medium including a stored program, wherein the program executes a method for processing image data in the embodiment.

[0144] According to an embodiment of the present invention, a processor is also provided for running a program, wherein the program executes the image data processing method of the embodiment during runtime.

[0145] According to another aspect of the present invention, a computer program product is also provided. The program product includes computer instructions that, when executed by a processor, implement the image data processing method described in the embodiments.

[0146] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0147] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0148] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0149] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0150] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0151] If the integrated unit is implemented as a software functional unit and sold or used as an independent functional component, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software functional component. This computer software functional component is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0152] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for processing image data, characterized in that, include: Collect image data of underground diseased areas; The image data is subjected to image enhancement and grayscale histogram equalization operations to obtain the preprocessed image data. Image recognition is performed on the preprocessed image data to obtain recognition results, wherein the recognition results include semantic category labels corresponding to multiple pixels in the image data, the semantic category labels are used to indicate the disease area category to which the corresponding pixel belongs, and the disease area category includes at least disease category and non-disease category; Based on the semantic category labels corresponding to the multiple pixels in the recognition results, the diseased areas and non-diseased areas in the image data are determined; The target location of the diseased area in the underground diseased area is subjected to three-dimensional visualization processing to obtain the detection result of the diseased area, wherein the detection result is used to represent the three-dimensional structural spatial features of the diseased area; The process involves performing three-dimensional visualization processing on the target location of the diseased area within the underground diseased area to obtain the detection result of the diseased area. This includes: extracting the geometric information of the diseased area, wherein the geometric information includes at least the coordinates of the center point of the diseased area and the size information of the diseased area; determining the target location of the diseased area within the underground diseased area based on the coordinates of the center point of the diseased area and the size information of the diseased area; acquiring image data at the target location; reconstructing the amplitude information of the image data at the target location at different depths based on an amplitude reconstruction algorithm, wherein the amplitude reconstruction algorithm is constructed based on inverse divergence imaging technology and offset technology; and determining the detection result of the diseased area based on the amplitude information.

2. The method according to claim 1, characterized in that, Image recognition is performed on the preprocessed image data to obtain recognition results, including: The preprocessed image data is input into the target defect detection model for image recognition to obtain the recognition result. The target defect detection model is obtained by training an initial defect detection model with image data samples of the defect area.

3. The method according to claim 2, characterized in that, The target defect detection model includes an encoder and a decoder. The encoder consists of a multi-channel convolutional feature selection module, a parallel multi-directional attention mechanism, and a residual compression module. The multi-channel convolutional feature selection module is used to extract features at different scales in the image data. The parallel multi-directional attention mechanism is used to enhance the perception of features at different scales. The residual compression module is used to compress the features at different scales. The decoder is used to reconstruct the image data based on the compressed features at different scales and generate the recognition result based on the reconstructed image data.

4. The method according to claim 3, characterized in that, The method further includes: Obtain image data samples of the diseased area; The image data samples are labeled to obtain the label data corresponding to the image data samples; The initial defect detection model is trained using the image data samples and the corresponding label data to obtain the target defect detection model.

5. The method according to claim 1, characterized in that, Based on the semantic category labels corresponding to the multiple pixels in the recognition results, the diseased and non-diseased regions in the image data are determined, including: The region corresponding to the pixel whose semantic category label is the disease category in the recognition result is determined as the disease region; The region corresponding to the pixel whose semantic category label in the recognition result is the non-disease category is determined as the non-disease region.

6. An image data processing apparatus, characterized in that, include: The acquisition unit is used to collect image data of underground diseased areas; The preprocessing unit is used to perform image enhancement and grayscale histogram equalization operations on the image data to obtain the preprocessed image data. The recognition unit is used to perform image recognition on the preprocessed image data to obtain a recognition result, wherein the recognition result includes semantic category labels corresponding to multiple pixels in the image data, the semantic category labels are used to indicate the disease area category to which the corresponding pixel belongs, and the disease area category includes at least a disease category and a non-disease category; The determining unit is used to determine the diseased area and non-diseased area in the image data based on the semantic category labels corresponding to the plurality of pixels in the recognition result; The processing unit is used to perform three-dimensional visualization processing on the target location of the diseased area in the underground diseased area to obtain the detection result of the diseased area, wherein the detection result is used to represent the three-dimensional structural spatial features of the diseased area; The processing unit is configured to perform three-dimensional visualization processing on the target location of the diseased area in the underground diseased area through the following steps to obtain the detection result of the diseased area: extracting the geometric information of the diseased area, wherein the geometric information includes at least the coordinates of the center point of the diseased area and the size information of the diseased area; determining the target location of the diseased area in the underground diseased area based on the coordinates of the center point of the diseased area and the size information of the diseased area; acquiring image data at the target location; reconstructing the amplitude information of the image data at the target location at different depths based on an amplitude reconstruction algorithm, wherein the amplitude reconstruction algorithm is constructed based on inverse divergence imaging technology and offset technology; and determining the detection result of the diseased area based on the amplitude information.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein when the program is run by a processor, it controls the device in which the storage medium is located to perform the method according to any one of claims 1 to 5.

8. A processor, characterized in that, The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 5 when it runs.

9. A computer program product, characterized in that, The computer program product includes computer instructions that, when executed by a processor, implement the method described in any one of claims 1 to 5.

Images