Intelligent karst groundwater exploration image processing system
By performing grayscale layering enhancement and edge repair on karst groundwater exploration images, the problems of difficulty in highlighting deep water-bearing anomalies and noise interference in karst areas in existing technologies have been solved, achieving efficient processing and accurate identification of karst groundwater exploration images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUIZHOU UNIV
- Filing Date
- 2026-03-31
- Publication Date
- 2026-05-29
AI Technical Summary
Existing karst groundwater exploration image processing systems neglect the physical characteristics of karst areas, making it difficult to highlight deep water-bearing anomalies. The weak signals from water-rich fissures cannot be captured, resulting in poor exploration image processing effects. Furthermore, there is amplified noise, producing jagged artifacts, leading to contour matching deviations and reducing the accuracy of karst damage identification.
By dividing the image grayscale into low-resistivity, transition, and high-resistivity regions, a membership function is designed for targeted enhancement. A dynamic haze rule library and a dark area enhancement power-state spreading function are used to improve grayscale stretching. Soxhlet gradient sharpening is used to accurately enhance edges. An apex constraint loss function is designed to accurately match irregular contours. A karst damage identification model is constructed to improve identification accuracy.
It improves the image processing effect of karst groundwater exploration, accurately captures weak signals, reduces noise interference, enhances the accuracy of karst damage identification, and realizes real-time on-site detection and feedback.
Smart Images
Figure CN121962779B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing, specifically to an intelligent image processing system for karst groundwater exploration. Background Technology
[0002] Karst groundwater exploration image processing systems refer to specialized technical systems that utilize computer vision technology to preprocess, enhance, extract features, and identify damage in raw exploration images obtained from high-density electrical resistivity tomography, ground-penetrating radar, and borehole imaging. This enables accurate detection and analysis of water-rich anomalies in karst caves and fissures. However, general karst groundwater exploration image processing systems often suffer from problems such as ignoring the physical properties of karst areas, failing to highlight deep water-bearing anomalies, and failing to capture weak signals from water-rich fissures, resulting in poor image processing performance. Furthermore, these systems often amplify noise, produce jagged artifacts at fissure edges, and experience contour matching deviations due to corner point misalignment, leading to poor accuracy in karst damage identification. Summary of the Invention
[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent image processing system for karst groundwater exploration. Addressing the problems of general karst groundwater exploration image processing systems neglecting the physical properties of karst areas, failing to highlight deep water-bearing anomalies, and failing to capture weak signals from water-rich fractures, resulting in poor image processing performance, this solution divides image grayscale into low-resistivity, transition, and high-resistivity zones based on the physical properties of the karst area. It also designs a membership function to achieve geologically guided directional enhancement. A dynamic blurring rule library is established based on the local window background mean, ensuring accurate adaptation of low-resistivity dark enhancement to the local background and preventing the obscuring of shallow weak anomalies. Furthermore, it employs a power-law enhancement method for dark areas. The expansion function increases the grayscale stretching slope in extremely low grayscale areas deep within karst caves and inside pipes, accurately capturing weak signals; thus improving the image processing effect of karst groundwater exploration. Addressing the issues of amplified noise, jagged artifacts at fissure edges, and contour matching deviations due to corner misalignment in general karst groundwater exploration image processing systems, leading to poor accuracy in karst damage identification, this solution uses Soxhlet gradient sharpening to precisely enhance edges, only enhancing the edge response of geological boundaries to avoid noise interference in non-boundary areas; a corner-end constraint loss function is designed to accurately match irregular contours and reduce the impact of slight offsets in the annotation box; thereby improving the accuracy of karst damage identification in karst groundwater exploration images.
[0004] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent karst groundwater exploration image processing system, including an image acquisition module, a geological body grayscale layer enhancement module, a geological feature stretching module, a karst boundary repair module, a damage location optimization module, a karst damage identification model establishment module, and a karst groundwater exploration image processing module;
[0005] The image acquisition module acquires historical karst exploration images, and after standardized preprocessing, constructs a karst exploration image set.
[0006] The geological body grayscale layering enhancement module enhances the karst exploration image set by adjusting the grayscale through a dynamic haze rule library and centroid method to achieve geological knowledge-guided enhancement.
[0007] The geological feature stretching module performs cropping, limiting equalization, and dark area enhancement power-law spreading on the karst exploration image after gray-scale layering of the geological body to achieve geological feature stretching.
[0008] The karst boundary repair module calculates the edge response of the karst exploration image after geological feature stretching, and repairs the karst boundary through kurtosis overlay and truncation processing to obtain the final enhanced image;
[0009] The damage localization optimization module is designed to incorporate a corner-peak constraint loss function that integrates overlap, center point, and corner-peak distance.
[0010] The karst damage identification model building module establishes a karst damage identification model based on the karst exploration image after karst boundary repair and the corner constraint loss function.
[0011] The karst groundwater exploration image processing module processes real-time karst exploration images based on a karst damage identification model.
[0012] Furthermore, the image acquisition module acquires historical karst exploration images, performs image standardization preprocessing and image annotation, and constructs a karst exploration image set. The image standardization preprocessing includes: scale unification, resampling to a uniform spatial resolution, and using center cropping to adjust all images to the standard size of the model input; grayscale normalization, converting color images into grayscale images; the image annotation uses an annotation tool to draw a rectangular box around the target on the image and specify the damage category label.
[0013] Furthermore, the geological body grayscale layer enhancement module, for the karst exploration image set, divides the grayscale of the karst exploration image into three geological layers—low resistivity zone, transition zone, and high resistivity zone—based on the physical characteristics of the karst area; and performs targeted enhancement on each layer, strengthening the darkness of the low resistivity zone, maintaining the grayscale of the transition zone, and increasing the brightness of the high resistivity zone, thereby achieving pre-enhancement that adds geological knowledge guidance to the image; specifically, it defines three attribution functions, establishes a dynamic haze rule library, modifies the attribution of each pixel, and finally uses the centroid method to calculate the determined output grayscale value.
[0014] Furthermore, the geological feature stretching module divides the karst exploration image set after layered enhancement into overlapping geological units, and performs a two-step sequential processing within each unit:
[0015] Step 1: Equalize the histogram within the cell by clipping and limiting it, while also preventing noise from being excessively amplified by clipping and limiting it.
[0016] Step two: Based on the initial equalization, a dark area enhancement power-state spreading function is used to redistribute the gray levels; the function introduces dark area enhancement sigmoid modulation to increase the slope of gray level stretching in the extremely low gray level region.
[0017] Furthermore, the karst boundary repair module uses the Soxhlet gradient sharpening property to enhance areas of abrupt grayscale changes in karst exploration images after geological feature stretching, resulting in a final enhanced image.
[0018] Furthermore, the damage localization optimization module designs a corner-peak constraint loss function that simultaneously considers overlapping areas, center point distance, and corner-peak distance to minimize the distance between the corresponding corners of the predicted bounding box and the ground truth bounding box.
[0019] Furthermore, the karst damage identification model building module takes the final enhanced image as input and constructs a karst damage identification model based on the vertex constraint loss function. The model adopts a backbone-neck-head architecture. The input is the final enhanced karst exploration image. The output is the category, confidence level, and bounding box coordinates of all karst exploration images in the image. The backbone network adopts a lightweight network constructed with depthwise separable convolution. The neck network adopts a weighted bidirectional feature pyramid network, which fuses and optimizes multi-scale features from the backbone network. The detection head adopts a decoupled head, using two independent branches to handle classification and regression tasks respectively.
[0020] Furthermore, the karst groundwater exploration image processing module acquires karst exploration images in real time, and after preprocessing, performs geological body grayscale layer enhancement, geological feature stretching, and karst boundary repair. Finally, the processed karst exploration images are sent to the karst damage recognition model, which outputs the type, confidence level, and bounding box of caves and fissures in real time, realizing on-site instant detection and feedback.
[0021] The beneficial effects achieved by the present invention using the above solution are as follows:
[0022] (1) In view of the problem that general karst groundwater exploration image processing systems ignore the physical characteristics of karst areas, make it difficult to highlight deep water-bearing anomalies, and fail to capture weak signals from water-rich fissures, thus resulting in poor exploration image processing effect, this scheme divides the image grayscale into low-resistivity, transition, and high-resistivity areas based on the physical properties of karst areas, and designs a degree function to achieve directional enhancement guided by geological knowledge; establishes a dynamic blurring rule library based on the local window background mean, so that the dark enhancement of low-resistivity areas can be accurately adapted to the local background, avoiding the masking of shallow weak anomalies; and improves the grayscale stretching slope of extremely low grayscale areas deep in caves and inside pipes by using a dark area enhancement power-state spreading function, so as to accurately capture weak signals; thereby improving the image processing effect of karst groundwater exploration.
[0023] (2) To address the problems of amplified noise in general karst groundwater exploration image processing systems, resulting in jagged artifacts at the edges of fissures and contour matching deviations due to corner misalignment, which leads to poor accuracy in karst damage identification, this solution uses Soxhlet gradient sharpening to precisely enhance the edges, only enhancing the edge response of geological boundaries to avoid noise interference in non-boundary areas; it designs a corner-peak constraint loss function to accurately match irregular contours and reduce the impact of slight offset of the annotation box; thereby improving the accuracy of karst damage identification in karst groundwater exploration images. Attached Figure Description
[0024] Figure 1 This is a flowchart illustrating an intelligent karst groundwater exploration image processing system provided by the present invention.
[0025] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0026] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0027] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0028] Example 1, see Figure 1 The present invention provides an intelligent karst groundwater exploration image processing system, including an image acquisition module, a geological body grayscale layer enhancement module, a geological feature stretching module, a karst boundary repair module, a damage location optimization module, a karst damage identification model establishment module, and a karst groundwater exploration image processing module;
[0029] The image acquisition module acquires historical karst exploration images, performs standardized preprocessing to construct a karst exploration image set, and sends the data to the geological body grayscale layering enhancement module.
[0030] The geological body grayscale layering enhancement module adjusts the grayscale of the karst exploration image set through a dynamic haze rule library and centroid method to achieve geological knowledge-guided enhancement; and sends the data to the geological feature stretching module.
[0031] The geological feature stretching module performs cropping, limiting equalization, and dark area enhancement power-law spreading on the karst exploration image after gray-scale layering of the geological body to achieve geological feature stretching; and sends the data to the karst boundary repair module.
[0032] The karst boundary repair module calculates the edge response of the karst exploration image after geological feature stretching, repairs the karst boundary through kurtosis overlay and truncation processing, and obtains the final enhanced image; and sends the data to the damage location optimization module.
[0033] The damage location optimization module is designed to integrate the overlap, center point and corner vertex distance into a corner vertex constraint loss function; and sends the data to the karst damage identification model building module.
[0034] The karst damage identification model building module establishes a karst damage identification model based on the karst exploration image after karst boundary repair and the corner constraint loss function; and sends the data to the karst groundwater exploration image processing module.
[0035] The karst groundwater exploration image processing module processes real-time karst exploration images based on a karst damage identification model.
[0036] Example 2, see Figure 1This embodiment is based on the above embodiment. The image acquisition module acquires historical karst exploration images, performs image standardization preprocessing and image annotation, and constructs a karst exploration image set. The historical karst exploration images include high-density electrical resistivity cross-section maps, ground-penetrating radar profile maps, and borehole imaging maps. Image standardization preprocessing includes: scale unification, resampling to a uniform spatial resolution (1 pixel = 0.1 meters), and using center cropping to adjust all images to the standard size of the model input (640x640 pixels); grayscale normalization, converting color images to grayscale images. The image annotation uses an annotation tool (LabelImg) to draw a rectangular box around the target on the image and specify the damage category label (cavity, fissure, dissolution conduit, weathering crust). All annotations must undergo a cross-review process: initial annotation by junior engineers → review by senior engineers → expert group to judge disputed samples to ensure the accuracy and consistency of the annotations.
[0037] Example 3, see Figure 1 This embodiment, based on the above embodiment, uses a geological body grayscale layer enhancement module to divide the grayscale of karst exploration images into three geological layers—low resistivity zone (water-rich target), transition zone (lithological interface), and high resistivity zone (intact surrounding rock)—according to the physical characteristics of the karst area. It then performs targeted enhancement on each layer, strengthening the darkness of the low resistivity zone to highlight water-bearing anomalies, maintaining the grayscale of the transition zone to preserve the lithological interface, and appropriately increasing the brightness of the high resistivity zone to distinguish the surrounding rock. This achieves a pre-enhancement layer guided by geological knowledge, highlighting water-bearing anomalies and preserving the lithological interface. The specific operation is as follows:
[0038] Define three membership functions , and , representing the degree to which the pixel grayscale value g belongs to the low, medium, and high resistivity regions, respectively. The function for classifying the pixel as belonging to the low resistivity region is expressed as: The transition zone affiliation function is expressed as: The high-resistivity region affiliation function is expressed as: g is the pixel grayscale value, and a, b, c, and d are function parameters that control the grayscale range of the low-resistivity region. a takes the first 20% of the histogram cumulative spread, b takes the first 35%, c takes the first 55%, and d takes the first 80%. , and These represent the confidence levels of a pixel belonging to the low-resistivity, medium-resistivity, and high-resistivity regions, respectively.
[0039] A dynamic haze rule base is established, and the affiliation degree of each pixel is modified. Finally, the centroid method is used to calculate the determined output grayscale value, which is expressed as: ; It is the output belonging degree after the i-th vague rule is activated. ; It represents the output grayscale center value of the corresponding rule; N is the total number of hazy rules; These are the grayscale values of the image pixels after layered enhancement; This is the average background value of a local window, with a window size of 5×5. It is the initial degree of belonging; It is a modulation function. ; It is the local background modulation coefficient, which is 0.2 in the low impedance region, 0.1 in the transition region, and 0.3 in the high impedance region; and These are the minimum and maximum gray values of the image, respectively.
[0040] Dynamic blurring rule base established: Rule 1: If a pixel belongs to a low-resistivity area, reduce its grayscale value; Rule 2: If a pixel belongs to a transition area, keep its grayscale value unchanged; Rule 3: If a pixel belongs to a high-resistivity area, enhance its grayscale value; The dynamic center value is automatically adjusted based on the local background mean to make the dark enhancement of water-rich targets more accurate; and an output grayscale center value is specified for each rule. ; ; , and These are the grayscale values after the pixel belongs to the low-resistivity region, the transition region, and the high-resistivity region, respectively. and These are empirical coefficients, all with values of (0.5, 0.8); m and n are neighborhood offset indices; It is the gray value of the image at the corresponding coordinates.
[0041] Example 4, see Figure 1 This embodiment is based on the above embodiment. For the karst exploration image set after layered enhancement, the geological feature stretching module divides the karst exploration image into overlapping geological units, and performs two-step sequential processing within each unit:
[0042] Step 1: Equalize the histogram within the unit by cropping and limiting it. This is to safely widen the overall grayscale dynamic range of the local area, initially improve the visual separability between the surrounding rock and the anomalous area, and at the same time prevent noise from being excessively amplified into a false anomaly by cropping and limiting it.
[0043] Step 2: Based on the initial equalization, a dark area enhancement power-state spreading function is used to redistribute the gray levels. The function introduces dark area enhancement sigmoid modulation, which can increase the slope of gray level stretching in the extremely low gray level region, thereby extremely finely depicting the gray level changes in the depth of the cave and inside the pipe, and has higher sensitivity to the weak signals of water-rich fissures.
[0044] The whole is represented as: ;in, F is a stretched karst exploration image showing geological features; F is a layered enhanced karst exploration image. It is the first output step; It is a step-by-step output; average number of pixels per block. for ; This is the block size, using 16×16; L is the number of gray levels within the block, 256; the maximum number of pixels allowed per gray level, CL, is expressed as: ; This is the clipping factor, ranging from 0.2 to 0.5; dark area enhancement power-state spreading mapping function. Represented as: ; ; It is the cumulative spread probability; It is a power factor. The larger the value, the stronger the stretching of low grayscale areas. The value ranges from 2 to 5. Histogram cropping and redistribution: For the histogram of each block, the part that exceeds CL is cropped, and the total amount of cropping is evenly added back to all grayscale levels to ensure that the total number of pixels remains unchanged and to avoid grayscale loss. It is a dark area enhancement function; This is the dark area threshold, set to 0.2L; Adjust the slope, set to 10-20; It is the intensity enhancement coefficient, with a value of 0.5 to 1.0, which controls the magnitude of the enhancement in dark area response.
[0045] By performing the above operations, this solution addresses the problems of general karst groundwater exploration image processing systems, which neglect the physical characteristics of karst areas, fail to highlight deep water-bearing anomalies, and cannot capture weak signals from water-rich fissures, resulting in poor image processing performance. Based on the physical properties of karst areas, this solution divides image grayscale into low-resistivity, transition, and high-resistivity zones, and designs a membership function to achieve geologically guided directional enhancement. A dynamic blurring rule library is established based on the local window background mean, ensuring accurate adaptation of low-resistivity dark enhancement to the local background and preventing shallow weak anomalies from being masked. Furthermore, by employing a dark area enhancement power-state spreading function, the grayscale stretching slope is increased in extremely low grayscale areas deep within karst caves and inside pipes, accurately capturing weak signals. This ultimately improves the image processing performance of karst groundwater exploration.
[0046] Example 5, see Figure 1 This embodiment is based on the above embodiment. After strong contrast stretching, the karst boundary repair module can cause the karst boundaries (cavity walls, fracture surfaces) in the image to become blurry or jagged. Therefore, for karst exploration images with stretched geological features, the Soxhlet gradient sharpening feature (extremely sensitive to areas with high changes in the first derivative of grayscale (i.e., edges)) is used to precisely enhance areas of abrupt grayscale changes, restoring the clear and sharp shape of cavity walls and fracture surfaces, making the geometric shape of the geological structure clearer; Soxhlet gradient sharpening output Represented as: ; ; and These are the convolution responses in the horizontal and vertical directions, respectively; the final kurtosis output is represented as: ; It ultimately enhances the image pixels; It is the kurtosis intensity coefficient, used to control the magnitude of edge enhancement, with a value ranging from 0.1 to 0.5; It is a truncation process to ensure that the pixel value is within the range of [0, 255].
[0047] Example 6, see Figure 1 This embodiment is based on the above embodiment. For noise annotation of karst caves and fissures in the image, the damage localization optimization module designs a corner-peak constraint loss function. It considers the overlapping area, the distance between the center point and the corner-peak distance, and explicitly minimizes the distance between the predicted box and the corresponding corner-peak of the real box. For karst damage, especially long strip fissures and irregular caves, it can accurately match the four corner points of the contour. Then, it accurately guides the network to adjust the predicted box so that the corner-peak is aligned with the corner point of the real geological damage, thus having stronger robustness to noise annotation.
[0048] Let the coordinates of the true bounding box A be... The coordinates of the predicted bounding box B are The image width and height are w and h; construct a vertex distance penalty term: calculate the normalized distance between the top left and bottom right vertices of the two bounding boxes. and , is represented as: ; Total loss Represented as: ; It is intersection, union, and ratio; It is the square of the Euclidean distance between the center points of frame A and frame B; Constraints are integrated for overlap, center point distance, and corner vertex distance; It is the distance penalty weight, with a value ranging from 0.5 to 1.
[0049] Example 7, see Figure 1This embodiment is based on the above embodiment. The karst damage identification model building module takes the final enhanced image as input and constructs a karst damage identification model based on the vertex constraint loss function. The model adopts a backbone-neck-head architecture. The input is the final enhanced karst exploration image. The output is the category, confidence level, and bounding box coordinates of all karst exploration images in the image. The backbone network adopts a lightweight network (MobileNetV3) constructed with depthwise separable convolution, which downsamples the image from shallow to deep layers and extracts feature maps of different scales (1 / 8, 1 / 16, 1 / 32 scales), corresponding to the features of small-scale fissures and large-scale caves. The neck network adopts a weighted bidirectional feature pyramid network, which integrates and optimizes multi-scale features from the backbone network. The contribution of each scale feature is dynamically adjusted through learnable weights to improve the multi-scale damage detection capability. The detection head adopts a decoupled head, using two independent branches to handle classification (damage category) and regression (bounding box coordinates) tasks respectively.
[0050] Regarding model training: The training objective of the model is to minimize the total loss, which consists of three parts: classification loss, using FocalLoss; regression loss, employing corner-end constraint loss to optimize the position of the bounding box; and targetability loss, determining whether the anchor box contains the target, using binary cross-entropy loss. The total loss is the weighted sum of these three losses, with the weights initially set at a ratio of 1.0:5.0:1.0, and then adjusted according to model performance. The final enhanced karst exploration image set is randomly divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set is used to monitor the training process, adjust hyperparameters, and select the best model, and the test set is used to finally evaluate the model's generalization performance. The AdamW optimizer is used for parameter tuning, and all trainable parameters in the model are updated in reverse based on the calculated total loss gradient. The learning rate is scheduled using a cosine annealing strategy. After each training epoch, the model performance (average precision and recall) is evaluated on the validation set. Training is stopped early when the validation set performance no longer improves over several consecutive epochs (improvement threshold of 0.001). The best-performing model is evaluated using a test set that has never been used in the training sequence. If the model performance meets the criteria (average precision reaches 85% at an IoU threshold of 0.5, and recall for the blemish category is not less than 80%), the model training is complete; otherwise, the initial parameters of the model are adjusted and retraining is performed.
[0051] By performing the above operations, this solution addresses the problems of amplified noise, jagged artifacts at fracture edges, and contour matching deviations caused by corner misalignment in general karst groundwater exploration image processing systems, which lead to poor accuracy in karst damage identification. This solution uses Soxhlet gradient sharpening to precisely enhance edges, only strengthening the edge response of geological boundaries to avoid noise interference in non-boundary areas. A corner-constrained loss function is designed to accurately match irregular contours and reduce the impact of slight offsets in the annotation box. This improves the accuracy of karst damage identification in karst groundwater exploration images.
[0052] Example 8, see Figure 1 This embodiment is based on the above embodiment. The karst groundwater exploration image processing module acquires karst exploration images in real time. After preprocessing, it performs grayscale layer enhancement of geological bodies, stretching of geological features, and repair of karst boundaries. Finally, the processed karst exploration images are sent to the karst damage recognition model, which outputs the type, confidence level, and bounding box of caves and fissures in real time, realizing on-site instant detection and feedback.
[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0054] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. An intelligent image processing system for karst groundwater exploration, characterized in that: The system includes an image acquisition module, a geological body grayscale layering enhancement module, a geological feature stretching module, a karst boundary repair module, a damage location optimization module, a karst damage identification model establishment module, and a karst groundwater exploration image processing module. The image acquisition module acquires historical karst exploration images, and after standardized preprocessing, constructs a karst exploration image set. The geological body grayscale layering enhancement module enhances the karst exploration image set by adjusting the grayscale through a dynamic haze rule library and centroid method to achieve geological knowledge-guided enhancement. The geological feature stretching module performs cropping, limiting equalization, and dark area enhancement power-law spreading on the karst exploration image after gray-scale layering of the geological body to achieve geological feature stretching. The karst boundary repair module calculates the edge response of the karst exploration image after geological feature stretching, and repairs the karst boundary through kurtosis overlay and truncation processing to obtain the final enhanced image; The damage localization optimization module is designed to incorporate a corner-peak constraint loss function that integrates overlap, center point, and corner-peak distance. The karst damage identification model building module establishes a karst damage identification model based on the karst exploration image after karst boundary repair and the corner constraint loss function. The karst groundwater exploration image processing module processes real-time karst exploration images based on a karst damage identification model. The damage localization optimization module designs a corner-peak constraint loss function that simultaneously considers overlapping areas, center point distance, and corner-peak distance. It explicitly minimizes the distance between the predicted bounding box and the corresponding corner-peak of the ground truth bounding box, thereby guiding the network to adjust the predicted bounding box so that the corner-peak aligns with the corner points of the actual geological damage. Let the coordinates of the true bounding box A be... The coordinates of the predicted bounding box B are The image width and height are w and h; construct a vertex distance penalty term: calculate the normalized distance between the top left and bottom right vertices of the two bounding boxes. and , is represented as: ; Total loss Represented as: ; It is intersection, union, and ratio; It is the square of the Euclidean distance between the center points of frame A and frame B; Constraints are integrated for overlap, center point distance, and corner vertex distance; It is the distance penalty weight.
2. The intelligent karst groundwater exploration image processing system according to claim 1, characterized in that: The geological body grayscale layer enhancement module divides the grayscale of karst exploration images into three geological layers—low resistivity zone, transition zone, and high resistivity zone—based on the physical properties of the karst area. It then performs targeted enhancement on each layer, strengthening the darkness of the low resistivity zone, maintaining the grayscale of the transition zone, and increasing the brightness of the high resistivity zone, thereby achieving pre-enhancement that adds geological knowledge guidance to the images. Specifically, three affiliation functions are defined, a dynamic haze rule base is established, the affiliation of each pixel is modified, and finally the centroid method is used to calculate the determined output grayscale value.
3. The intelligent karst groundwater exploration image processing system according to claim 1, characterized in that: The geological feature stretching module divides the enhanced karst exploration image set into overlapping geological units, and performs a two-step sequential processing within each unit: Step 1: Equalize the histogram within the cell by clipping and limiting it, while also preventing noise from being excessively amplified by clipping and limiting it. Step two: Based on the initial equalization, a dark area enhancement power-state spreading function is used to redistribute the gray levels; the function introduces dark area enhancement sigmoid modulation to increase the slope of gray level stretching in the extremely low gray level region.
4. The intelligent karst groundwater exploration image processing system according to claim 1, characterized in that: The karst boundary repair module uses Soxhlet gradient sharpening to enhance areas of abrupt grayscale changes in karst exploration images after geological feature stretching, resulting in a final enhanced image.
5. The intelligent karst groundwater exploration image processing system according to claim 1, characterized in that: The karst damage identification model building module takes the final enhanced image as input and constructs a karst damage identification model based on the vertex constraint loss function; the model adopts a backbone-neck-head architecture; The input is the final enhanced karst exploration image; the output is the category, confidence level, and bounding box coordinates of all karst exploration images in the image; the backbone network is a lightweight network constructed using depthwise separable convolutions; the neck network is a weighted bidirectional feature pyramid network that fuses and optimizes multi-scale features from the backbone network. The detection head uses a decoupled head, employing two independent branches to handle classification and regression tasks respectively.
6. The intelligent karst groundwater exploration image processing system according to claim 1, characterized in that: The image acquisition module acquires historical karst exploration images, performs image standardization preprocessing and image annotation, and constructs a karst exploration image set. Image normalization preprocessing includes: scale unification, resampling to a uniform spatial resolution, and center cropping to adjust all images to the standard size of the model input; Grayscale normalization converts a color image into a grayscale image; image annotation involves using an annotation tool to draw a rectangular box around the target on the image and specifying a damage category label.
7. The intelligent karst groundwater exploration image processing system according to claim 1, characterized in that: The karst groundwater exploration image processing module acquires karst exploration images in real time. After preprocessing, it performs grayscale layer enhancement of geological bodies, stretching of geological features, and repair of karst boundaries. Finally, the processed karst exploration images are sent to the karst damage recognition model, which outputs the type, confidence level, and bounding box of caves and fissures in real time, enabling real-time on-site detection and feedback.