Mining area ground fracture measuring system based on improved RetinaNet

Abstract

The invention discloses a mining area ground fracture measurement system based on improved RetinaNet, and belongs to the field of mining area ground fracture measurement. The system is composed of a data acquisition module, a data processing module, a data calculation module, a data storage module and a data visualization module. An unmanned aerial vehicle is adopted to carry a high-resolution camera to collect ground fracture data, an improved RetinaNet model is used for accurately identifying boundary, form, position and other information of a ground fracture according to a preprocessed image, and the identified ground fracture is measured, calculated and imported into a data storage module; meanwhile, a model generated through deep learning training is stored, and model fine adjustment and optimization are carried out subsequently; and finally, the recognition and measurement results are displayed to the user in the forms of a three-dimensional model, a chart and a data report through a graphical user interface. And a user comprehensively masters the ground fracture detection and analysis process. Through cooperative work of the five modules, automatic identification and accurate measurement of the ground fissures of the mining area are realized, and the method is excellent in processing complex terrains and multi-scale fissure characteristics.

Description

Technical Field

[0001] This invention relates to a mining area ground fissure measurement system based on an improved RetinaNet, belonging to the field of mining area ground fissure measurement. Background Technology

[0002] In mine management and geological monitoring, the identification and measurement of ground fissures are crucial. Ground fissures not only may indicate potential geological disasters, but also have a direct impact on the safe production and maintenance of the mine. Traditional ground fissure detection methods usually rely on manual exploration and simple image analysis techniques. These methods are not only time-consuming and labor-intensive, but also struggle to guarantee accuracy and comprehensiveness in complex terrain and harsh environments.

[0003] In recent years, with the continuous development of the mining industry and the increase in mining depth, the monitoring of ground fissures in mining areas has become more complex and important. The formation of ground fissures in mining areas is influenced by various factors, including geological structure, mining activities, and changes in the natural environment. The appearance of ground fissures not only leads to problems such as surface deformation and groundwater pollution in mining areas, but may also trigger serious geological disasters such as landslides and collapses, posing a significant threat to the safe operation of mining areas. Therefore, timely and accurate identification and measurement of ground fissures are of great significance for preventing geological disasters and ensuring the safety of mining areas.

[0004] Traditional methods for monitoring ground fissures mainly include manual inspection, ground radar detection, and satellite remote sensing. However, these methods have many limitations. For example, manual inspection is time-consuming, costly, and easily affected by human factors; ground radar detection, although highly accurate, has limited coverage and is complex to operate; satellite remote sensing technology, although with a wide coverage, has insufficient resolution and is difficult to capture small fissures (Zhang Junyang, Wang Kun, Zhao Tongbin, et al. Research status and development of UAV remote sensing observation of surface subsidence and fissures in mining areas [J]. Coal Science and Technology, 2024, 52(S2):435-444.DOI:10.13199 / j.cnki.cst.2023-0438.; Wang Yadong, Lu Liuyan. Research on UAV image acquisition and processing methods in deformation monitoring of coal mine goaf [J]. Industrial Automation, 2024, 50(S2):72-75.). Therefore, there is an urgent need for an efficient, accurate, and automated method for measuring ground fissures to overcome the shortcomings of traditional technologies.

[0005] With the rapid development of computer vision and deep learning technologies, image-based automated crack identification technology has gradually become a research hotspot. Utilizing drones for high-resolution image acquisition, combined with advanced deep learning algorithms, automated identification and measurement of ground cracks can be achieved. This method not only improves monitoring efficiency and accuracy but also enables large-scale, blind-spot-free monitoring in harsh environments and complex terrains. Furthermore, through continuous algorithm optimization and improved computing power, automated crack identification technology has demonstrated outstanding performance in handling multi-scale and multi-type ground cracks, gradually becoming the mainstream trend in ground crack monitoring in mining areas. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a mining area ground fissure measurement system based on an improved RetinaNet model. The improved RetinaNet model performs exceptionally well in handling complex terrain and multi-scale fissure features. It integrates five modules—data acquisition, processing, calculation, storage, and visualization—to achieve efficient detection and accurate measurement of mining area ground fissures.

[0007] This invention uses an improved RetinaNet model as its core, combining five modules working collaboratively: data acquisition, processing, computation, storage, and visualization. It utilizes a drone equipped with a high-resolution camera to acquire high-resolution images of the mining area, addressing the limitations of traditional methods in terms of coverage or accuracy. The acquired images are preprocessed (including CLAHE enhancement, ADnet denoising, and Gaussian filtering) to improve image quality and provide reliable data for subsequent identification. Based on the improved RetinaNet model (adding dilated convolutions to the ResNet backbone to enhance the receptive field while preserving fine-grained features), it achieves automated identification of ground fissures, using Focal Loss to address class imbalance and improve recognition accuracy. Geometric features of the identified ground fissures are measured (through bounding box coordinate extraction, binarization, skeletonization, and pixel-to-actual-length conversion) to obtain key parameters such as fissure length. The original data, processing results, and trained model are stored to support subsequent model optimization. Results are displayed through visualization methods such as 3D models and charts, combined with an interactive interface to allow users to fully control the detection process. Ultimately, this achieves efficient and accurate identification and measurement of ground fissures in mining areas, providing technical support for geological disaster prevention and mining safety management.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A ground fissure measurement system based on an improved RetinaNet includes a data acquisition module, a data processing module, a data calculation module, a data storage module, and a data visualization module, which together enable accurate identification and measurement of ground fissures.

[0010] The data acquisition module uses a drone equipped with a high-resolution camera to collect data on ground fissures.

[0011] In the data processing module, the image is preprocessed to improve image quality and remove unnecessary interference, and the preprocessed image is further analyzed and labeled.

[0012] The data calculation module uses the improved RetinaNet model to accurately identify the boundaries, shape, and location of ground fissures in the processed image, and measures and calculates the identified ground fissures.

[0013] The data storage module is used to store the processed and calculated data. This module not only saves the original data and processing results, but also saves the model generated by deep learning training, so as to carry out subsequent model fine-tuning and optimization.

[0014] The data visualization module presents the identification and measurement results to the user in the form of 3D models, charts, and data reports through a graphical user interface. Users can view detailed analysis results, adjust parameters, and operate the system through the graphical interface to fully control the ground fissure detection and analysis process.

[0015] Through the seamless integration and collaboration of the above five modules, the system enables comprehensive measurement and analysis of ground fissures in mining areas, providing an efficient, accurate, and intuitive geological safety monitoring solution.

[0016] The components included in each module are as follows:

[0017] Data acquisition module: A DJI M350RTK drone, equipped with a high-resolution camera, is used to acquire ground images of the mining area.

[0018] Data processing module: This includes a computer using a Lenovo Ultra9285K processor and an RTK5080 graphics card. It runs image processing algorithms (such as the CLAHE algorithm, ADnet Gaussian noise removal algorithm, and Gaussian filtering algorithm) to perform image preprocessing, feature extraction, and other operations. Data acquisition module connects to data processing module: The drone transmits acquired image data to the computer device in the data processing module in real time via wireless transmission (such as Wi-Fi, 4G / 5G), or imports the stored image data into the computer device via a data cable after the flight.

[0019] Data computation module: Employs a dedicated GPU server to run the improved RetinaNet model, enabling computational tasks such as ground fissure identification and length measurement. The data processing module and the data computation module interact via an internal data bus (such as the PCIe bus on a computer motherboard). Pre-processed image data from the data processing module is transmitted to the data computation module for the improved RetinaNet model to perform identification and measurement calculations.

[0020] Data storage module: Includes a database server and hard disk array, used to store raw image data, preprocessed data, annotation information, trained models, etc. Data computation module connects to the database server via a network (such as a local area network) to transmit recognition results, measurement data, and trained models to the storage devices of the data storage module for saving.

[0021] Data visualization module: A computer terminal or mobile device equipped with a display for showing 3D models, charts and data reports, while supporting user interaction.

[0022] The computer terminal or mobile device of the data visualization module accesses the database server of the data storage module through the network, reads the required ground fissure identification and measurement data, and uses it to generate and display 3D models, charts, etc.

[0023] Overall Collaborative Relationship: The five modules form a complete data flow closed loop through the above connection method. The data acquisition module obtains the raw data, which is then preprocessed by the data processing module. The core identification and measurement are completed by the data calculation module, and the results are stored in the data storage module. Finally, the data visualization module presents the results to the user, realizing the full-process identification and measurement of ground fissures in the mining area.

[0024] This invention provides a method for measuring ground fissures in mining areas based on an improved RetinaNet, comprising the following steps:

[0025] Step s1: Use a drone equipped with a high-resolution camera to fly along a predetermined path and take ground images of a specific area of ​​the mining area. Build an image database from the collected images and label the images. Divide the images into training set, test set and validation set according to the size of the data.

[0026] The specific process is as follows: The drone cruises at a set altitude and speed, while the camera takes pictures at certain time intervals or distances. The shooting interval can be flexibly adjusted in the system according to factors such as the actual size of the mining area and the density of ground fissures. Generally, it is set to take one image every 5-10 meters of flight to ensure the continuity and completeness of the acquired images and to comprehensively capture ground fissure information. The images are transmitted in real time to a ground base station via wireless transmission (such as Wi-Fi, 4G / 5G), and then the ground base station transmits the data to the data processing center. Alternatively, after the drone's flight mission is completed, the image data in the storage device can be imported into the data processing center via a data cable, thereby constructing a mining area ground fissure image database. Once the database has accumulated a certain amount of image data, the images are divided into training, testing, and validation sets in a 6:2:2 ratio according to the data volume. If the data volume is small (less than 1000 images), the data is directly divided after manual checking of the labeling accuracy; if the data volume is large (more than 1000 images), 10% of the data is randomly sampled for quality inspection to ensure that the labeling quality meets the standards before the data is divided. The partitioning process uses a random sampling algorithm to ensure that the data distribution of each subset is uniform, covering ground fissure images of different types, sizes and characteristics, in order to meet the needs of subsequent training, testing and validation of the RetinaNet model, and improve the model's accuracy and generalization ability in identifying and measuring ground fissures in mining areas.

[0027] Step s2: The data processing module is connected to the data acquisition module and is used to preprocess, extract features and augment data of ground fissure image data. Based on the processed ground fissure image data, the improved RetinaNet is used to identify and extract ground fissures, and at the same time, a three-dimensional model of a specific area with ground fissure annotation information is generated.

[0028] The specific process is as follows: First, image preprocessing is performed, using the CLAHE algorithm for histogram equalization to enhance image contrast and highlight the outline of the ground fissures. An ADnet-based Gaussian noise removal algorithm, combined with attention mechanisms, dense connections, and multi-scale feature extraction, removes noise while preserving edge features. Gaussian filtering further reduces noise interference and improves image quality. Second, key features such as texture, shape, and edges of the ground fissures in the processed image are extracted. Next, the dataset is expanded by rotating, scaling, flipping, and adjusting brightness of the image to enhance the model's adaptability to ground fissures under different angles, scales, and lighting conditions. Finally, based on the processed image data, image stitching and 3D reconstruction techniques are used to generate a 3D model of a specific region with ground fissure annotation information, providing a geometric reference for subsequent spatial distribution analysis and measurement of ground fissures.

[0029] Step s3: The data calculation module is connected to the data processing module to perform geometric feature measurements on the identified ground fissures and calculate the length of the fissures.

[0030] Step s4: The data storage module is used to store the processed data, establish a database of ground fissure images in the mining area, and store the trained model. Subsequent model fine-tuning is then possible.

[0031] Step s5: Based on the data acquisition module, data processing module, data calculation module, and data storage module, the data visualization module is used to generate three-dimensional models and charts of ground fissure identification and measurement results; it includes a graphical user interface and allows users to operate and control the system.

[0032] Furthermore, in step s1, the data acquisition module uses a drone equipped with a high-resolution camera to fly in a specific area of ​​the mining area along a pre-designed path. During flight, the drone captures real-time ground images, and all captured images are stored in a database. The images required for the experiment are labeled with ground fissures; then, the data is divided into training, testing, and validation sets according to a certain ratio. The training set is used for model learning, the testing set is used for periodic evaluation during model training, and the validation set is used for final model performance evaluation and validation.

[0033] Furthermore, in step s2, the data processing module first preprocesses the ground fissure image data, including data augmentation, denoising, and filtering:

[0034] Histogram equalization using the CLAHE algorithm effectively enhances image contrast, making object outlines clearer and more distinct.

[0035] The Gaussian noise removal algorithm based on ADnet is used. By introducing an attention mechanism to focus on important regions in the image, and combining dense connections and multi-scale feature extraction, noise can be effectively removed while preserving the edge features of the image, thereby enhancing the accuracy and stability of target detection.

[0036] The Gaussian filtering algorithm is used to process the image to reduce noise interference, thereby providing a higher quality image for subsequent target detection.

[0037] Furthermore, in step s2, the improved RetinaNet model is pre-trained using sample images of ground fissures to identify their boundaries and morphological information. During this training process, the algorithm learns the characteristics of various ground fissures (trained using images from a database). Cracks are detected, bounding boxes are drawn at their locations on the images, and labels are added for subsequent analysis. Through multiple iterations and optimizations, the model improves its accuracy in identifying ground fissures in different environments and complex backgrounds. Finally, the fully trained model can accurately detect and analyze ground fissures in practical applications, providing reliable data support for subsequent data storage, data visualization, and security assessment.

[0038] Furthermore, the improved RetinaNet model in step S2 mainly consists of a ResNet module, an FPN feature pyramid, a classification subnetwork, and a bounding box regression subnetwork. The backbone network, composed of ResNet and FPN, is used to extract features from the target. The classification subnetwork and bounding box regression subnetwork are responsible for target classification and location regression on the feature maps output by the FPN, respectively. Specifically, anchor boxes are generated on each feature map, and a pre-set intersection-over-union (IoU) threshold is used to match the anchor boxes with the target. Then, feature maps with different anchor boxes are input into the classification subnetwork and the bounding box regression subnetwork to predict the probability of target occurrence and determine its specific location. Simultaneously, the algorithm proposes a Focal loss function, effectively alleviating the problem of low detection accuracy caused by imbalanced sample classes.

[0039] The backbone network is a residual neural network, ResNet. The ResNet residual network outputs a portion of the original data directly to the next layer through shortcut connections in the residual blocks, balancing the contradiction between linear and nonlinear transformations. It solves the problems of gradient descent and accuracy saturation without adding extra network parameters.

[0040] In image processing, low-level features contain less semantic information but have more precise target location, while high-level features contain rich semantic information but relatively coarse location information. The feature pyramid model combines high-level semantic features with low-level location features through top-down operations and lateral connections, allowing the image to retain high resolution while expanding the receptive field, thereby enhancing the feature extraction capability of the neural network.

[0041] The fused feature layers are fed into the classification sub-network and the bounding box sub-network to obtain the location of the ground fissures. The classification sub-network first performs convolution operations, and then uses the ReLU activation function after each convolution. The Softmax function is used to obtain the confidence of each anchor box. During training, anchor boxes that reach the confidence threshold are selected and Focal Loss is used to calculate the classification loss. The network is then optimized through learning. The classification sub-network and the bounding box sub-network do not share parameters, but parameter sharing is achieved between the feature layers.

[0042] The ReLU activation function effectively avoids the vanishing gradient problem, is computationally simple, and can improve model training efficiency. Its specific formula is:

[0043] f(x) = max(0,x)

[0044] Where: if x is greater than 0, then f(x) = x; if x is less than or equal to 0, then f(x) = 0.

[0045] The softmax function transforms a real-valued vector into a probability distribution, with output values ​​ranging from [0,1], and the sum of all output values ​​being 1. It is well-suited for the output layer of classification problems. Its formula is:

[0046] Where: z i It is the i-th element in the input vector. It sums all input elements after applying exponents to ensure that the sum of all outputs is 1.

[0047] Focal Loss addresses the class imbalance and easy / difficult sample imbalance issues in object detection by controlling the weights of positive / negative and easy / difficult samples.

[0048] FL(p t )=-α(1-p t ) γ logp t

[0049]

[0050] Where, p t This represents the score of the predicted bounding box classification, α represents the parameter used to control the balance between positive and negative samples, and γ represents the modulation factor parameter, which reduces the proportion of easily classified samples in the learning process. Focal Loss not only reduces the contribution of simple samples to the classification loss function, but also expands the loss range of highly difficult incorrect samples, achieving a balance between the contributions of positive and negative samples in the classification process and improving the detection performance of single-stage detection models.

[0051] Furthermore, the improvement to the RetinaNet model in step S2 is specifically as follows: the backbone network of RetinaNet is modified in a targeted manner, and dilated convolutions are added on the basis of ResNet. While extracting semantic information, fine-grained low-level detail features are preserved. That is, holes are added to ordinary convolution kernels to increase the receptive field of the original convolution kernels. The modified convolution blocks aim to increase the receptive field while maintaining the spatial resolution of the feature maps.

[0052] Further, in step S3, the data calculation module obtains the coordinates of each bounding box of the crack output by the model. These coordinates can be used for further processing and measurement of the crack. The extracted crack region is binarized to generate a binary image. Image skeletonization technology is used to convert the binary crack image into a skeleton image, where the pixels in the skeleton image represent the centerline of the crack. The crack centerline pixels in the skeleton image are traversed, and the total length of these pixels is calculated. Based on the image resolution, the pixel length is converted into the actual length unit (cm).

[0053] Furthermore, in step S4, the data storage module not only stores the processed data but also establishes a comprehensive database of ground fissure images in the mining area. This database contains high-quality images that have been preprocessed, labeled, and classified to facilitate subsequent analysis and research. In addition, the data storage module also stores models trained through deep learning, including the initial training model and versions optimized through multiple iterations. In future practical applications, these pre-trained models can be directly used for ground fissure detection, and the models can be fine-tuned based on new data and needs to further improve their detection accuracy and robustness. In this way, the data storage module not only provides a reliable data foundation but also ensures the sustainable optimization and application of the models.

[0054] Furthermore, in step S5, the data visualization module establishes a graphical user interface and a 3D model and charts of the cracks, including a 3D display unit for displaying the 3D model and charts of the generated ground crack identification and measurement results. This unit can not only render the 3D structure of the ground cracks in real time, providing an intuitive visualization effect, but also display detailed measurement data and statistical analysis results. Through the interactive interface, users can rotate, zoom, and pan the 3D model to observe the specific morphology and location of the cracks in detail. Simultaneously, the 3D display unit also supports various chart formats, such as line charts, bar charts, and heat maps, intuitively displaying the distribution, trends, and key data indicators of the cracks, helping users to more comprehensively understand and analyze the geological conditions of the mining area.

[0055] The graphical user interface (GUI) displays the results of ground fissure identification and measurement via computer or mobile device, allowing users to adjust parameters and operate the system, providing an intuitive and easy-to-use interactive platform. Users can view detailed 3D models, charts, and data reports of ground fissures to understand the specific results of identification and measurement. The interface features clear navigation and operation guidance, enabling users to easily browse different analysis views and report content. Furthermore, users can adjust various parameters, such as adjusting detection sensitivity, selecting different display modes, and setting warning thresholds, thereby personalizing the system's functionality and performance. Through the GUI, users can also perform system operations, such as starting or stopping detection tasks, importing or exporting data, and viewing historical records, gaining comprehensive control over the ground fissure detection and analysis process.

[0056] The beneficial effects of this invention are:

[0057] (1) The present invention proposes a mining area ground fissure measurement system based on an improved RetinaNet, which integrates five major modules: data acquisition, processing, calculation, storage and visualization, and successfully realizes rapid identification, accurate measurement and precise evaluation of mining area ground fissures;

[0058] (2) Images of the mining area were collected by drones equipped with high-resolution cameras, and key feature information of ground fissures was extracted through preprocessing and annotation;

[0059] (3) The improved RetinaNet algorithm is trained in the data computing module to identify and extract the boundary and morphological information of ground fissures;

[0060] (4) The processing results and models are stored in the data storage module, providing a basis for subsequent model optimization; through the data visualization module, users can intuitively view the three-dimensional model and analysis results, and fully control the process of ground fissure detection and analysis; this integrated system not only improves the efficiency and accuracy of ground fissure detection, but also provides solid technical support for mine safety management. Attached Figure Description

[0061] Figure 1 Here is the flowchart for the CLAHE algorithm;

[0062] Figure 2 Here is a flowchart of the ADnet algorithm;

[0063] Figure 3 This is an architecture diagram of the mining area ground fissure measurement system based on the improved RetinaNet of this invention;

[0064] Figure 4 This refers to the information on ground fissures in the mining area obtained by the acquisition module in this embodiment of the invention.

[0065] Figure 5 This is a schematic diagram of the RetinaNet model of the present invention.

[0066] Figure 6 This is a comparative schematic diagram of the improved structure of the RetinaNet model of this invention. Detailed Implementation

[0067] The present invention will be further illustrated by the following embodiments, but is not limited to the following embodiments.

[0068] Example 1:

[0069] To make the overall process and technology of this invention clearer, embodiments of the invention are described below. Furthermore, to provide the public with a deeper understanding of this invention, some details of the specific implementation process are shown below.

[0070] It should be understood that the specific embodiments described herein are for illustrative purposes only and should not be construed as limiting the invention.

[0071] See Figure 3This invention provides a mining area ground fissure measurement system based on an improved RetinaNet, which consists of a data acquisition module, a data processing module, a data calculation module, a data storage module, and a data visualization module connected in sequence. The data acquisition module uses a drone equipped with a high-resolution camera to collect ground fissure data. The pre-processed images are then used to accurately identify the boundaries, morphology, and location of the ground fissures using an improved RetinaNet model. The identified ground fissures are measured and calculated, and the data is imported into the data storage module. Simultaneously, the model generated through deep learning training is saved for subsequent model fine-tuning and optimization. Finally, the system connects to the data visualization module, which displays the identification and measurement results to the user in the form of 3D models, charts, and data reports through a graphical user interface. The user can view detailed analysis results, adjust parameters, and operate the system through the graphical interface, gaining full control over the ground fissure detection and analysis process. Through the collaborative work of these five modules, automated identification and accurate measurement of ground fissures in mining areas are achieved.

[0072] The method for measurement using the above system specifically includes the following steps:

[0073] Step s1: The data acquisition module uses a drone equipped with a high-resolution camera to fly along a predetermined path and capture ground images of a specific area of ​​the mining area. The acquired images are used to build an image database, and the images are labeled. Based on the amount of data, the images are divided into training set, test set and validation set according to a fixed ratio.

[0074] Step s2: The data processing module is connected to the data acquisition module and is used to preprocess, extract features, and augment data of the ground fissure image data. Based on the processed ground fissure image data, the improved RetinaNet is used to identify and extract ground fissures, and at the same time, a three-dimensional model of a specific area with ground fissure annotation information is generated.

[0075] Step s3: The data calculation module is connected to the data processing module to perform geometric feature measurements on the identified ground fissures and calculate the length of the fissures.

[0076] Step s4: The data storage module is used to store the processed data, establish a database of ground fissure images in the mining area, and store the trained model. Subsequent model fine-tuning is then possible.

[0077] Step s5: Based on the data acquisition module, data processing module, data calculation module, and data storage module, the data visualization module is used to generate three-dimensional models and charts of ground fissure identification and measurement results; it includes a graphical user interface and allows users to operate and control the system.

[0078] In this embodiment, see Figure 4In step S1, the data acquisition module uses a DJI Matrice 350RTK drone equipped with a high-resolution camera to collect ground fissure information in a specific area above the working face of a mining area, following a pre-designed path. Each image has a resolution of 5280×39956 pixels, and the captured images are stored in a database. All images required for the experiment are labeled with ground fissures. The data is then divided into a training set, a test set, and a validation set in a 6:2:2 ratio. The training set is used for model learning, the test set is used for periodic evaluation during model training, and the validation set is used for final model performance evaluation and validation.

[0079] In this embodiment, the data processing module in step S2 first preprocesses the ground fissure image data, using the CLAHE algorithm for histogram equalization, which effectively enhances image contrast and makes object outlines clearer and more distinct; it then uses an ADnet-based Gaussian noise removal algorithm, which introduces an attention mechanism to focus on important regions in the image, and combines dense connections and multi-scale feature extraction to effectively remove noise while preserving the image's edge features, thereby enhancing the accuracy and stability of target detection; finally, it uses a Gaussian filtering algorithm to process the image to reduce noise interference, thus providing higher quality image assurance for subsequent target detection.

[0080] In this embodiment, the improved RetinaNet model in step S2 is pre-trained using sample images of ground fissures to identify their boundaries and morphological information. During this training process, the algorithm learns the characteristics of various ground fissures, detects the fissures, draws bounding boxes at the fissure locations on the images, and adds labels to the fissures for subsequent analysis. Through multiple iterations and optimizations, the model improves its accuracy in identifying ground fissures in different environments and complex backgrounds. Finally, the fully trained model can accurately detect and analyze ground fissures in practical applications, providing reliable data support for subsequent data storage, data visualization, and security assessment.

[0081] In this embodiment, see Figure 5The improved RetinaNet model in step S2 mainly consists of a ResNet backbone network, an FPN feature pyramid, a classification subnetwork, and a bounding box regression subnetwork. The backbone network is composed of ResNet and FPN to extract features from the target. The classification subnetwork and bounding box regression subnetwork are responsible for target classification and location regression on the feature maps output by the FPN, respectively. Specifically, anchor boxes are generated on each feature map, and a pre-set intersection-over-union (IoU) threshold is used to match the anchor boxes with the target. Then, the feature maps with different anchor boxes are input into the classification subnetwork and the bounding box regression subnetwork to predict the probability of target occurrence and determine its specific location. Simultaneously, the algorithm proposes a Focalloss loss function, effectively alleviating the problem of low detection accuracy caused by imbalanced sample classes.

[0082] In this embodiment, the system utilizes an improved RetinaNet model (introducing dilated convolutions into the ResNet backbone) to expand the receptive field while maintaining high-resolution feature maps, significantly enhancing its ability to identify ground fissures in multi-scale, complex backgrounds. In this embodiment, after training, the model's loss value decreased to 0.284, achieving a recognition accuracy of 72%, indicating good detection performance and generalization ability in real mining environments. Simultaneously, it achieves automated and high-precision measurement.

[0083] The backbone network is a residual neural network, ResNet. The ResNet residual network outputs a portion of the original data directly to the next layer through shortcut connections in the residual blocks, balancing the contradiction between linear and nonlinear transformations. It solves the problems of gradient descent and accuracy saturation without adding extra network parameters.

[0084] In image processing, low-level features contain less semantic information but have more precise target location, while high-level features contain rich semantic information but relatively coarse location information. The feature pyramid model combines high-level semantic features with low-level location features through top-down operations and lateral connections, allowing the image to retain high resolution while expanding the receptive field, thereby enhancing the feature extraction capability of the neural network.

[0085] The fused feature layers are fed into the classification sub-network and the bounding box sub-network to obtain the location of the ground fissures. The classification sub-network first performs convolution operations, and then uses the ReLU activation function after each convolution. The Softmax function is used to obtain the confidence of each anchor box. During training, anchor boxes that reach the confidence threshold are selected and Focal Loss is used to calculate the classification loss. The network is then optimized through learning. The classification sub-network and the bounding box sub-network do not share parameters, but parameter sharing is achieved between the feature layers.

[0086] The ReLU activation function effectively avoids the vanishing gradient problem, is computationally simple, and can improve model training efficiency. Its specific formula is:

[0087] f(x) = max(0, x)

[0088] Where: if x is greater than 0, then f(x) = x; if x is less than or equal to 0, then f(x) = 0.

[0089] The softmax function transforms a real-valued vector into a probability distribution, with output values ​​ranging from [0,1], and the sum of all output values ​​being 1. It is well-suited for the output layer of classification problems. Its formula is:

[0090] Where: z i It is the i-th element in the input vector. It sums all input elements after applying exponents to ensure that the sum of all outputs is 1.

[0091] Focal Loss addresses the class imbalance and easy / difficult sample imbalance issues in object detection by controlling the weights of positive / negative and easy / difficult samples.

[0092] FL(p t )=-α(1-p t ) γ logp t

[0093]

[0094] Where pt is the score of the predicted box classification, α represents the parameter used to control the balance between positive and negative samples. In this embodiment, α = 0.25, γ = 2, and γ represents the modulation factor parameter, which is a parameter to reduce the proportion of easily classified samples in the learning process. Focal Loss not only reduces the contribution of simple samples to the classification loss function, but also expands the loss range of highly difficult incorrect samples, achieving a balance between the contributions of positive and negative samples in the classification process and improving the detection performance of the single-stage detection model.

[0095] In this embodiment, see Figure 6The specific improvements to the RetinaNet model in step S2 are as follows: The RetinaNet backbone network is specifically modified by adding dilated convolutions to the ResNet architecture. The first convolutional block in stages 4 and 5, which previously had a stride of 2, is replaced with a dilated convolutional layer. This ensures that the spatial size of the feature map remains downsampled eight times relative to the original image from stage 4 onwards. While extracting semantic information, fine-grained low-level details are preserved by adding dilation to ordinary convolutional kernels to increase their receptive field. The modified convolutional blocks aim to increase the receptive field while maintaining the spatial resolution of the feature map.

[0096] In this embodiment, in step S3, the data calculation module obtains the coordinates of each bounding box of the crack output by the model. These coordinates can be used for further processing and measurement of the crack. The extracted crack region is binarized to generate a binary image. Image skeletonization technology is used to convert the binary crack image into a skeleton image, where the pixels in the skeleton image represent the centerline of the crack. The crack centerline pixels in the skeleton image are traversed, and the total length of these pixels is calculated. Based on the image resolution, the pixel length is converted to centimeters.

[0097] In this embodiment, the improved RetinaNet model iterates continuously during training, and the final loss value of the model in this embodiment is 0.284, with a recognition accuracy of 72%. Therefore, the improved RetinaNet-based system and method for measuring ground fissures in mining areas proposed in this invention has excellent recognition performance.

[0098] In this embodiment, in step S4, the data storage module not only stores the processed data but also establishes a comprehensive database of ground fissure images in the mining area. This database contains high-quality images that have been preprocessed, labeled, and classified to facilitate subsequent analysis and research. Furthermore, the data storage module also stores models trained through deep learning, including the initial training model and versions optimized through multiple iterations. In future practical applications, these pre-trained models can be directly used for ground fissure detection, and the models can be fine-tuned based on new data and requirements to further improve their detection accuracy and robustness. In this way, the data storage module not only provides a reliable data foundation but also ensures the sustainable optimization and application of the models.

[0099] In this embodiment, in step S5, the data visualization module establishes a graphical user interface and a 3D model and charts of the cracks, including a 3D display unit for displaying the 3D model and charts of the generated ground crack identification and measurement results. This unit can not only render the 3D structure of the ground cracks in real time, providing an intuitive visualization effect, but also display detailed measurement data and statistical analysis results. Through the interactive interface, users can rotate, zoom, and pan the 3D model to observe the specific shape and location of the cracks in detail. Simultaneously, the 3D display unit also supports various chart formats, such as line charts, bar charts, and heat maps, intuitively displaying the distribution, changing trends, and key data indicators of the cracks, helping users to more comprehensively understand and analyze the geological conditions of the mining area.

[0100] The graphical user interface (GUI) displays the results of ground fissure identification and measurement via computer or mobile device, allowing users to adjust parameters and operate the system, providing an intuitive and easy-to-use interactive platform. Users can view detailed 3D models, charts, and data reports of ground fissures to understand the specific results of identification and measurement. The interface features clear navigation and operation guidance, enabling users to easily browse different analysis views and report content. Furthermore, users can adjust various parameters, such as adjusting detection sensitivity, selecting different display modes, and setting warning thresholds, thereby personalizing the system's functionality and performance. Through the GUI, users can also perform system operations, such as starting or stopping detection tasks, importing or exporting data, and viewing historical records, gaining comprehensive control over the ground fissure detection and analysis process.

[0101] In this embodiment, the mining area ground fissure measurement system based on the improved RetinaNet consists of a data acquisition module, a data processing module, a data calculation module, a data storage module, and a data visualization module, which together achieve accurate identification and measurement of ground fissures above a mining working face. First, the data acquisition module uses a DJI Matrice 350RTK equipped with a high-resolution camera to collect ground fissure data. In the data processing module, the images undergo histogram equalization using the CLAHE algorithm, and Gaussian filtering is applied to improve image quality and remove unnecessary interference. The preprocessed images are then further analyzed and labeled. Next, the data calculation module uses the improved RetinaNet model to accurately identify the boundaries, morphology, and location of the ground fissures in the processed images, and performs measurement calculations on the identified fissures. The final model loss value is 0.284, achieving an identification accuracy of 72%. The data storage module stores the original data and processing results, as well as the trained improved RetinaNet model. Finally, the data visualization module displays the identification and measurement results in the form of 3D models, charts, and data reports through a graphical user interface, providing comprehensive control over the ground fissure detection and analysis process. Through the seamless integration and collaboration of these five modules, the system enables comprehensive measurement and analysis of ground fissures in mining areas, providing an efficient, accurate, and intuitive geological safety monitoring solution.

[0102] The above embodiments are only used to illustrate the specific process and architecture of the present invention, and are not intended to limit it; with reference to the detailed description of this embodiment, any person skilled in the art should understand that equivalent substitutions can still be made to some or all of the technical features of the technical solutions in the foregoing embodiments, and these adjustments or substitutions will not cause the substance of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and still fall within the protection scope of the technical solutions of the present invention.

Claims

1. A mine area ground fissure measuring system based on improved RetinaNet, characterized in that: The data acquisition module, the data processing module, the data calculation module, the data storage module and the data visualization module are sequentially connected to form a whole. The data acquisition module uses a high-resolution camera carried by a UAV to collect ground fissure data. The preprocessed image is used to accurately identify the boundary, shape and position of the ground fissure by using an improved RetinaNet model. The identified ground fissure is measured and calculated, and is imported into the data storage module. Meanwhile, the model generated by deep learning training is saved for subsequent model fine-tuning and optimization. Finally, the data visualization module is connected, and the identification and measurement results are displayed to the user in the form of three-dimensional models, charts and data reports through a graphical user interface. The user can view detailed analysis results, adjust parameters and operate the system through the graphical interface, and fully control the ground fissure detection and analysis process. The automatic identification and accurate measurement of the ground fissure in the mining area are realized through the cooperative work of the five modules.

2. The improved RetinaNet-based mine area ground fissure measuring system according to claim 1, characterized in that: The data acquisition module uses a DJI M350RTK UAV carrying a high-resolution camera to fly and collect ground images in a specific area of the mining area according to a preset path. The image resolution is not less than 5280x39956 pixels. The collected images are directly stored in the database and labeled for ground fissures, and then divided into a training set, a test set and a validation set in a ratio of 6:2:

2.

3. The improved RetinaNet-based mine area ground fissure measuring system according to claim 1, characterized in that: The data processing module includes a computer using a Lenovo Ultra9 285K processor and an RTK5080 graphics card for running image processing algorithms to complete image preprocessing and feature extraction operations. The data acquisition module is connected to the data processing module: the UAV transmits the collected image data to the computer device of the data processing module in real time through wireless transmission, or imports the stored image data into the computer device through a data line after the flight is completed.

4. The improved RetinaNet-based mine area ground fissure measuring system according to claim 1, characterized in that: The data calculation module uses an independent GPU server to run the improved RetinaNet model to realize the calculation task of ground fissure identification and length measurement. The data processing module and the data calculation module interact with each other through an internal data bus. The preprocessed image data of the data processing module is transmitted to the data calculation module for identification and measurement calculation by the improved RetinaNet model.

5. The improved RetinaNet-based mine area ground fissure measuring system according to claim 1, characterized in that: The data storage module includes a database server and a hard disk array for storing original image data, preprocessed data, labeling information and trained models. The data calculation module is connected to the database server through a network, and the identification results, measurement data and trained models are transmitted to the storage device of the data storage module for storage.

6. The improved RetinaNet-based mine area ground fissure measuring system according to claim 1, characterized in that: The data visualization module is a computer terminal or a mobile device equipped with a display for displaying three-dimensional models, charts and data reports, and supporting user interaction. The computer terminal or mobile device of the data visualization module accesses the database server of the data storage module through a network to read the required ground fissure identification and measurement data for generating three-dimensional models, charts and displaying.

7. A mine area ground fissure measuring method based on improved RetinaNet, adopting the mine area ground fissure measuring system based on improved RetinaNet in any one of claims 1-6. The method comprises the following steps: Step s1, use a drone to carry a high-resolution camera, fly according to the predetermined path, shoot the ground image of the specific area of the mining area, build an image database from the collected image, and label the image, according to the data size, the pictures are divided into training set, test set and validation set according to the fixed ratio; Step s2, the data processing module is connected with the data acquisition module, which is used for pre-processing, feature extraction and data enhancement of the ground fissure image data, and the improved RetinaNet is used for ground fissure recognition and extraction based on the processed ground fissure image data, and a specific area three-dimensional model with ground fissure annotation information is generated; Step s3, the data calculation module is connected with the data processing module, the geometric characteristics of the recognized ground fissure are measured, and the length of the fissure is calculated; Step s4, the data storage module is used for storing the processed data, establishing the mining area ground fissure image database, and storing the generated model, which can be fine-tuned in the future; Step s5, based on the data acquisition module, data processing module, data calculation module and data storage module, the data visualization module is used to generate the three-dimensional model and chart of the ground fissure recognition and measurement results; Including a graphical user interface, and allowing users to operate and control the system.

8. The mine area ground fissure measuring method based on improved RetinaNet according to claim 7, characterized in that: The specific process of step s1 is that the unmanned aerial vehicle cruises at a set height and speed, at the same time, the camera shoots according to the set time interval or distance interval, the shooting interval is flexibly adjusted in the system according to the actual mining area scale, ground fissure density factors, and is set to shoot one image every 5-10 meters of flight to ensure the coherence and integrity of the obtained image and fully capture the ground fissure information; through wireless transmission, the data is transmitted to the ground base station in real time, and then the ground base station transmits the data to the data processing center, or after the unmanned aerial vehicle completes the flight task, the image data in the storage device is imported into the data processing center through the data line, thereby constructing the mining area ground fissure image database; when the database accumulates a certain amount of image data, the pictures are divided into training set, test set and validation set according to the ratio of 6:2:2 according to the data size.

9. The mine area ground fissure measuring method based on improved RetinaNet according to claim 7, characterized in that: The specific process of step s2 is as follows: first, the image is pre-processed, the CLAHE algorithm is used for histogram equalization, the image contrast is enhanced to highlight the ground fissure contour, the ADnet-based Gaussian noise removal algorithm is used, combined with attention mechanism, dense connection and multi-scale feature extraction to remove noise and retain edge features, and Gaussian filter is used to further reduce noise interference and improve image quality; secondly, the texture, shape and edge key features of the ground fissure in the processed image are extracted; thirdly, the image is rotated, scaled, flipped and brightness adjusted to expand the data set and enhance the model's adaptability to ground fissures under different angles, scales and lighting conditions; finally, based on the processed image data, combined with image stitching and three-dimensional reconstruction technology, a specific area three-dimensional model with ground fissure annotation information is generated, which provides geometric reference for subsequent spatial distribution analysis and measurement of ground fissures.

10. The mine area ground fissure measuring method based on improved RetinaNet according to claim 9, characterized in that: The improved RetinaNet model in the step S2 mainly consists of a ResNet module, a FPN feature pyramid, a classification subnetwork and a bounding box regression subnetwork; wherein the ResNet and the FPN jointly constitute a backbone network to realize feature extraction of a target; the classification subnetwork and the bounding box regression subnetwork are respectively responsible for target classification and position regression of feature maps output by the FPN, specifically, anchor boxes are generated on each feature map, a pre-set intersection-over-union threshold is used to match the anchor boxes and the target, then the feature maps with different anchor boxes are input into the classification subnetwork and the bounding box regression subnetwork to predict a target occurrence probability and determine a specific position of the target; meanwhile, the algorithm proposes a Focal loss loss function, which effectively alleviates the problem of low detection accuracy caused by sample category imbalance.

Images