Surrounding rock blast hole trace detection method and system based on ResUNet-GhostNetV3 network

Abstract

The invention discloses a surrounding rock blast hole trace detection method and system based on a ResUNet-GhostNetV3 network, and relates to the technical field of tunnel construction, and the method comprises the steps: obtaining a surrounding rock image data set; preprocessing the surrounding rock image data set, performing binary feature extraction on the surrounding rock image data set by using a ResUNet network, and constructing a training data set; a blast hole trace detection model is constructed based on a GhostNetV3 convolutional neural network; training and verifying the blast hole trace detection model through the training data set to obtain a final blast hole trace detection model; and detecting surrounding rock blast hole traces through the final blast hole trace detection model. The method can quickly detect and count blast hole residues, effectively evaluate the blasting effect, and is beneficial to dynamically adjusting the blasting design and construction according to data, improving the construction quality and reducing the construction cost.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel construction, and more specifically to a method and system for detecting blast hole marks in surrounding rock based on a ResUNet-GhostNetV3 network. Background Art

[0002] Drilling and blasting is one of the main construction methods for modern mountain tunnels. Smooth blasting technology is widely used because it can well control the quality and stability of tunnel formation. The setting of reasonable parameters for blasting is complex and needs to take into account the entire mining geological conditions. Any method for determining the setting of reasonable parameters for blasting should consider the influence of geological spatial variability to ensure blasting effect and construction safety.

[0003] However, the setting of tunnel blasting parameters depends on the subjective experience of construction workers. Tunnel over-excavation and under-excavation and safety accidents often occur due to improper setting of blasting parameters. As one of the important indicators for evaluating blasting quality, the blast hole residual rate and blast hole trace rate of the tunnel face can reflect the over-excavation and under-excavation of the surrounding rock and the lithology differences.

[0004] At present, the hole marks left after blasting are mainly statistically evaluated by manual marking, and the detection efficiency and accuracy are low. This will lead to deviations between the blasting effect evaluation and the actual blasting situation on site, resulting in errors in the blasting effect evaluation results, affecting the subsequent optimization and adjustment of blasting design and construction.

[0005] With the rapid development of deep learning, detection methods based on image technology have begun to be applied in the field of engineering construction. Due to its automated parameter learning ability, it effectively overcomes the defects of poor generalization ability of traditional image processing methods and can adapt to complex and changeable engineering application scenarios. For example, Zhang Wanzhi studied the blasthole recognition technology of rock tunnel face based on FasterR-cnn to assist in optimizing blasting parameters. KIM et al. proposed to use the VGG model to classify the surrounding rock images of the face, but the recognition effect and accuracy were not ideal. Xian Qingyu et al. proposed a method based on the DenseNet network to distinguish the interference of the surrounding rock images of the face, and obtained the automatic extraction of the surrounding rock classification feature parameters. Liu Quan improved the YOLOv7 target detection model by introducing the Simple Attention Module (SimAm) and changing the loss function (Enhanced Intersection Over Union Loss, EIoU) to solve the problem of rapid and accurate detection of blasthole images in underground mine blasting charges, and optimized the blasthole detection feature network. In order to realize the automatic loading of explosives, Zhang Xingchang studied the application of deep learning-based target detection algorithm in blasthole detection. The blast hole images collected at the tunnel construction site were used to create a data set, and the Faster R-CNN and YOLOv5 algorithms were used for training and testing, and the advantages and disadvantages of the two algorithms in various performances were compared. However, the detection of blast hole marks on the tunnel wall is affected by factors such as irregular rock texture and fracture interference, and is heavily dependent on manual marking, resulting in problems such as high detection difficulty and low efficiency.

[0006] Therefore, how to propose a method and system for detecting blasthole marks in surrounding rock based on the ResUNet-GhostNetV3 network, quickly detect and count blasthole marks, effectively evaluate the blasting effect, and dynamically adjust the blasting design and construction according to the data to improve the construction quality and reduce the construction cost is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0007] In view of this, the present invention provides a method and system for detecting blasthole marks in surrounding rock based on the ResUNet-GhostNetV3 network, which can quickly detect and count blasthole marks and effectively evaluate the blasting effect. At the same time, it is beneficial to dynamically adjust the blasting design and construction according to the data, improve the construction quality and reduce the construction cost.

[0008] In order to achieve the above object, the present invention adopts the following technical scheme: a surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network, comprising:

[0009] Acquire surrounding rock image dataset;

[0010] Preprocessing the surrounding rock image dataset, and using the ResUNet network to extract binary features from the surrounding rock image dataset to construct a training dataset;

[0011] Construct a blast hole trace detection model based on GhostNetV3 convolutional neural network;

[0012] The blast hole mark detection model is trained and verified by using the training data set to obtain a final blast hole mark detection model;

[0013] The blast hole traces in the surrounding rock are detected through the final blast hole trace detection model.

[0014] Preferably, preprocessing the surrounding rock image dataset includes:

[0015] The surrounding rock image dataset is annotated using an annotation tool to complete data annotation.

[0016] Preferably, the ResUNet network includes an encoder and a decoder with jump connections;

[0017] The encoder is used to extract abstract representations, and the path of the encoder includes: repeated convolution layers, batch normalization layers, and maximum pooling layers in sequence;

[0018] The decoder is used to restore the original resolution, and the path of the decoder includes a transposed convolution layer and an upsampling layer.

[0019] Preferably, a reparameterization method is used to introduce linear parallel branches into depth-wise convolution and 1x1 convolution when training the blast hole mark detection model, thereby improving the learning ability of the model.

[0020] Preferably, the GhostNetV3 convolutional neural network includes multiple Ghost modules;

[0021] The Ghost module is used to reduce the number of parameters and computational cost of ordinary convolution.

[0022] Preferably, the process of reducing the number of parameters and computational cost by the Ghost module includes: generating a current feature Y, performing a deep convolution on the current feature Y, and generating an original feature Y′;

[0023] The original feature Y′ and the fixed feature are connected along the channel dimension to obtain the final output Y″, which is expressed as:

[0024] Y = X * W;

[0025] is the convolution kernel, X is the input feature, c in and c outRespectively represent the input and output channel dimensions, k is the kernel size, and * represents the convolution operation;

[0026] Y′=X*W p ;

[0027] Y″=Cat(Y′,X*W c );

[0028] Among them, W p and W c They represent the main convolution operation and the parameters in the operation respectively, and Cat represents the connection operation.

[0029] Preferably, the surrounding rock blasthole trace detection system based on the ResUNet-GhostNetV3 network includes:

[0030] A data set acquisition module, used to acquire surrounding rock image data sets;

[0031] A processing module, used for preprocessing the surrounding rock image data set, and using a ResUNet network to extract binary features from the surrounding rock image data set to construct a training data set;

[0032] Model building module, used to build a blasthole trace detection model based on GhostNetV3 convolutional neural network;

[0033] A training module, used for training and verifying the blast hole mark detection model through the training data set to obtain a final blast hole mark detection model;

[0034] The detection module is used to detect the blasthole marks in the surrounding rock through the final blasthole mark detection model.

[0035] It can be seen from the above technical scheme that compared with the prior art, the present invention discloses a method and system for detecting blast hole marks in surrounding rock based on the ResUNet-GhostNetV3 network, including: obtaining a surrounding rock image dataset; preprocessing the surrounding rock image dataset, and using the ResUNet network to perform binary feature extraction on the surrounding rock image dataset to construct a training dataset; constructing a blast hole mark detection model based on the GhostNetV3 convolutional neural network; training and verifying the blast hole mark detection model through the training dataset to obtain the final blast hole mark detection model; detecting surrounding rock blast hole marks through the final blast hole mark detection model.

[0036] The present invention is based on the prior art. In terms of noise robustness, the surface shadows, textures and cracks in the surrounding rock image are the main factors affecting the detection accuracy. The verification of the traditional blast hole trace detection algorithm relies on manual marking, which cannot completely avoid noise factors, resulting in low detection efficiency and delays in construction progress. In order to quickly and effectively detect the blast hole trace rate after tunnel blasting and improve the efficiency of tunnel construction quality inspection, the present invention proposes a ResUNet-GhostNetV3 network based on image features to study the detection of tunnel surrounding rock blast hole traces. Combined with experiments, the model detection accuracy, feature extraction accuracy and noise robustness are compared and analyzed. (1) In terms of feature extraction, the blast hole trace detection algorithm based on the GhostNetV3 network is based on the manual marking results. The feature extraction ability of the ResUNet network for blast hole traces is compared and analyzed, and the model parameters can be adaptively trained and optimized and high-level image features can be extracted. Experimental results show that the ResUNet-GhostNetV3 network can effectively avoid interference factors such as shadows, joints and cracks that are similar to blast hole traces, and is superior to traditional methods in feature extraction efficiency and noise robustness. (2) In terms of detection accuracy, the pixel accuracy of the method of the present invention is 87.9%, which is close to the accuracy of the detection method of manual marking, and can meet the needs of detecting the trace rate of surrounding rock blastholes in engineering. At the same time, the research results of the present invention also lay a technical foundation for the realization of comprehensive automation and refined blasting parameter control in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0038] Figure 1 A schematic flow chart of a method for detecting blasthole marks in surrounding rock based on the ResUNet-GhostNetV3 network provided in the present invention.

[0039] Figure 2 A schematic diagram of the ResUNet-GhostNetV3 network structure provided in an embodiment of the present invention.

[0040] Figure 3 A schematic diagram of the ResUNet network structure provided in an embodiment of the present invention.

[0041] Figure 4 A schematic diagram of the GhostNetV3 network structure provided in an embodiment of the present invention.

[0042] FIG. 5( a ) is a schematic diagram of a segmented image of rock surface shadow interference provided by an embodiment of the present invention.

[0043] FIG5( b ) is a schematic diagram of an artificially marked image of rock surface shadow interference provided by an embodiment of the present invention.

[0044] FIG5( c ) is a schematic diagram of a binary image output by a ResUNet network with rock surface shadow interference provided by an embodiment of the present invention.

[0045] FIG6( a ) is a schematic diagram of a segmented image of rock mass joint and fissure interference provided by an embodiment of the present invention.

[0046] FIG6( b ) is a schematic diagram of an artificially marked image of rock mass joint and fissure interference provided by an embodiment of the present invention.

[0047] FIG6( c ) is a schematic diagram of a binary image output by a ResUNet network of rock joint and fissure interference provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] The rate of blasthole marks is an important basis for evaluating the blasting quality of drilling and blasting, and is usually counted based on image recording and manual visual inspection. In order to overcome the problems of low precision and low efficiency of manual annotation of traditional image processing methods, the embodiment of the present invention discloses a method and system for detecting blasthole marks in surrounding rocks based on the ResUNet-GhostNetV3 network, and outputs the accuracy of blasthole mark detection. First, the surrounding rock image is taken using the ResUNet network for binary feature extraction and a blasthole mark data set is constructed. Secondly, the GhostNetV3 network is trained using the data set. Finally, a comparative analysis is performed in combination with the manually labeled data set, and the detection results of blasthole marks are output. The experimental results show that the detection accuracy of the ResUNet binary feature set is similar to that of the GhostNetV3 network on the manually labeled verification set, indicating that the ResUNet-GhostNetV3 network can quickly and effectively detect the blasthole marks in the surrounding rocks, and through auxiliary evaluation of the blasting effect, it is conducive to dynamically adjusting the blasting design and construction according to the data, improving the construction quality and reducing the construction cost.

[0050] In a specific embodiment of the present invention, a surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network is provided. Figure 1 As shown, including:

[0051] Acquire surrounding rock image dataset;

[0052] Preprocessing the surrounding rock image dataset, and using the ResUNet network to extract binary features from the surrounding rock image dataset to construct a training dataset;

[0053] Construct a blast hole trace detection model based on GhostNetV3 convolutional neural network;

[0054] The blast hole mark detection model is trained and verified by using the training data set to obtain a final blast hole mark detection model;

[0055] The blast hole traces in the surrounding rock are detected through the final blast hole trace detection model.

[0056] The embodiment of the present invention avoids the problem of low efficiency of manually marking blasthole marks in traditional methods, introduces the ResUNet network to automatically extract blasthole marks, and generates a data set. The GhostNetV3 network is trained based on the data set to obtain its optimal hyperparameters. Through verification of the test set, the accuracy of the ResUNet-GhostNetV3 network for blasthole mark detection is output. And through comparative analysis with the data set of manually labeled blasthole marks, the ability of the ResUNet network to extract features of blasthole marks is verified.

[0057] Specifically, preprocessing the surrounding rock image data set includes:

[0058] The surrounding rock image dataset is annotated using an annotation tool to complete data annotation.

[0059] Specifically, the data set of the embodiment of the present invention includes 180 24-bit RGB images with a size of 6000×4000 pixels. After the tunnel blasting, the technicians took the surrounding rock images with a handheld digital camera perpendicular to the surface of the tunnel wall. The data set was annotated manually using an annotation tool, using straight line segments to annotate along the symmetry axis of the blasthole marks, and the line width was uniformly set to 10 pixels.

[0060] Another method uses the ResUNet network to extract binary features from the surrounding rock images and construct a blasthole trace dataset.

[0061] Specifically, the ResUNet network is used for automatic recognition to generate binary images and report errors. In order to expand the data set, a Python script is used to capture images on the original image with a 512x512 sliding window. After image enhancement technology, a training data set of about 18,000 images of size 512x512 is obtained.

[0062] Specifically, the ResUNet network includes an encoder and a decoder with jump connections;

[0063] The encoder is used to extract abstract representations, and the path of the encoder includes: repeated convolution layers, batch normalization layers, and maximum pooling layers in sequence;

[0064] The decoder is used to restore the original resolution, and the path of the decoder includes a transposed convolution layer and an upsampling layer.

[0065] Specifically, UNet is characterized by an encoder-decoder structure with skip connections, which allows low-level details from the encoder to be fused with high-level semantic features from the decoder. The contracted encoder path consists of repeated convolutions, batch normalization, and max pooling layers to extract abstract representations. The expanded decoder path includes transposed convolutions and upsampling layers to restore the original resolution. Skip connections connect encoder and decoder features at each level in series to prevent spatial information loss and achieve accurate boundary positioning. In an embodiment of the present invention, Figure 3 As shown, the kernel size and number of filters of the architecture are set to 3 in all convolutional layers, suitable for image segmentation. The network starts with 64 filters in the first layer and gradually increases in deeper layers to promote hierarchical learning. After 4 decoding layers, the output passes through the last fully connected convolutional layer, in which the kernel size is adjusted according to the number of categories in the mask to meet the needs of the specific task. The output is then activated using a function that varies with the number of labels, ensuring customized results for each task.

[0066] ResUNet extends UNet by incorporating residual connections into the encoder path. Each encoder block contains residual units with identity mapping, which facilitates gradient flow and enables training deeper networks. The decoder mirrors the encoder by upsampling and connecting skip connections. Residual blocks help reuse and refine features across layers.

[0067] Specifically, a re-parameterization method is used to introduce linear parallel branches into depth-wise convolution and 1x1 convolution when training the blast hole mark detection model to improve the learning ability of the model.

[0068] Specifically, the GhostNetV3 convolutional neural network includes multiple Ghost modules;

[0069] The Ghost module is used to reduce the number of parameters and computational cost of ordinary convolution.

[0070] The GhostNetV3 adopted in the embodiment of the present invention not only inherits the advantages of the GhostNetV2 model in terms of low parameter count and detection speed, but also achieves significant performance improvement through innovative training strategies.

[0071] Specifically, the process of reducing the number of parameters and the computational cost by the Ghost module includes: generating a cost feature Y, performing a deep convolution on the cost feature Y, and generating an original feature Y′;

[0072] The original feature Y′ and the fixed feature are connected along the channel dimension to obtain the final output Y″, which is expressed as:

[0073] Y = X * W;

[0074] is the convolution kernel, X is the input feature, c in and c out Respectively represent the input and output channel dimensions, k is the kernel size, and * represents the convolution operation;

[0075] Y′=X*W p ;

[0076] Y″=Cat(Y′,X*W c );

[0077] Among them, W p and W c They represent the main convolution operation and the parameters in the operation respectively, and Cat represents the connection operation.

[0078] Specifically, Figure 4 As shown in Figure 1, GhostNet is a state-of-the-art compact module designed for efficient inference. The compact module is designed for edge device applications, with fast inference speed but average performance. However, the current training strategy for compact models borrows from the training strategy of traditional models, ignoring the difference in model capacity, which may affect the performance of compact models. Considering the impact of different training components, a strong training strategy for compact models is introduced.

[0079] In ordinary convolution, the output feature Y is obtained by Y = X * W, where is the convolution kernel, and X is the input feature. in and c out denote the input and output channel dimensions, respectively. k is the kernel size, and * denotes the convolution operation. The Ghost module reduces the number of parameters and computational cost of ordinary convolution in two steps.

[0080] First, the cost feature Y is generated, whose channel dimension is smaller than the original feature Y′. Then, deep convolution is performed on the inherent feature to generate feature Y′. The final output Y″ is obtained by connecting the inherent feature and the fixed feature along the channel dimension, which can be expressed as:

[0081] Y′=X*W p ;

[0082] Y″=Cat(Y′,X*W c);

[0083] Among them, W p and W c Represent the main convolution operation and the parameters in the operation respectively. "Cat" represents the connection operation. The entire GhostNet model is composed of multiple Ghost modules stacked together.

[0084] Specifically, depth-wise convolution and 1x1 convolution are common components in compact model architectures. Inspired by the successful experience of traditional model training, a reparameterization method is adopted for these two compact modules to obtain better performance. When training the compact model, linear parallel branches are introduced to the depth-wise convolution and 1x1 convolution. These additional parallel branches can be reparameterized after training and do not incur additional costs during inference.

[0085] like Figure 2 As shown in the figure, GhostNetV3 adopts a reparameterization method to improve the learning ability of the model by using depth-separable convolution and point group convolution without introducing additional parameters and computational complexity in the inference stage. Reparameterization operations are used for 3x3 depth convolution and 1x1 convolution modules, and it is found that the direct connection (identity) branch has a positive effect on improving model performance.

[0086] The embodiment of the present invention develops the GhostNetV3 convolutional neural network based on the Pytorch framework, uses the Adam optimizer as the parameter optimization algorithm, and the learning rate is 10 -4 The training process adopts a small batch stochastic gradient descent strategy, with a batch size of 16 in each iteration. After 100,000 iterations, the network loss function gradually stabilizes, the network model tends to converge, and the training ends.

[0087] like Figure 2 As shown, the surrounding rock blast hole mark detection method based on the ResUNet-GhostNetV3 network provided by the embodiment of the present invention first avoids the problems of low efficiency and poor accuracy of the traditional method of manually marking blast hole marks, and considers introducing the ResUNet network to automatically extract the features of the blast hole marks, obtain the binary image of the blast hole marks and cache it, and construct a training data set. Secondly, the GhostNetV3 network is trained and verified multiple times using the training set, and its hyperparameter model with the highest accuracy is retained. Finally, the model is used to verify the test set to obtain the detection results of the blast hole mark detection. A comparative analysis is performed in combination with the data set of manually marked blast hole mark positions to verify the accuracy of the ResUNet network in extracting the features of the blast hole marks.

[0088] In order to fully verify the accuracy of the tunnel blasting hole residue image detection method, the comparative analysis is completed from three aspects: detection accuracy, feature extraction, and noise robustness. In this embodiment of the present invention, the dataset of manually annotated images and the two datasets after ResUNet image processing are used as mutually verified datasets, and the accuracy of GhostNetV3 detection is used for comparative analysis.

[0089] Specifically, model detection accuracy comparison:

[0090] In order to verify that GhostNetV3 has achieved better results in detecting blast hole trace features, under the same experimental conditions, a binary image dataset generated by ResUNet was used, and then trained on VGG16, GoogleNet, InceptionV3, ResNet34, and DenseNet121. The results of different algorithms for malicious code behavior were evaluated, and the results are shown in Table 1.

[0091] Table 1 Model accuracy comparison data table / (m 3 )

[0092]

[0093] As shown in Table 1, using the GhostNetV3 network, the accuracy of the binary image data set generated based on ResUNet reached 87.9%. At the same time, the GhostNetV3 proposed in the present invention is much better than other common detection models in other evaluation indicators, and can provide a reliable basis for the calculation of the surrounding rock blasthole trace rate.

[0094] Specifically, feature extraction:

[0095] In order to verify the ability of ResUNet network to extract the features of blast hole marks after smooth blasting in tunnels, the GhostNetV3 detection network was compared with the manually labeled feature dataset, and the results are shown in Table 2.

[0096] Table 2 Model accuracy comparison data table / (m 3 )

[0097]

[0098]

[0099] As shown in Table 2, using the GhostNetV3 network, the accuracy based on the manually annotated data set is 91.82%, and the detection rate based on the binary image data set generated by ResUNet reaches 87.9%. At the same time, the ResUNet-GhostNetV3 proposed in the present invention is close to the GhostNetV3 detection method based on the manually annotated data set in terms of evaluation indicators, proving that ResUNet can provide a reliable basis for the feature extraction of blasthole traces in surrounding rock.

[0100] Specifically, noise robustness:

[0101] According to the categories of interference factors in the surrounding rock images, the test images are considered to be divided into two categories to compare the noise robustness performance of the ResUNet-GhostNetV3 network, as shown in Table 3. Among them, the first category of images is rock surface shadow interference; the second category of images is rock joint crack interference.

[0102] (1) The first type of image is the rock surface shadow interference

[0103] like Figure 5(a)-Figure 5(c) As shown in Figure 5(c), for the analysis of the first type of image noise samples, the image is blurred, the blast hole marks are faded, and some colors are not much different from the rock background, making it difficult to distinguish with the naked eye. In the results of the ResUNet-GhostNetV3 network detection, the detection results of most objects are consistent with the manually labeled data set, and some blast hole marks are discontinuous, which can meet the requirements of blast hole mark rate calculation, as shown in Figure 5(c). Among them, it should be emphasized that the cache results of the GhostNetV3 network in the ResUNet network to obtain blast hole marks are basically consistent with the detection results of the manually labeled data set, and have excellent robustness and real-time performance.

[0104] (2) The second type of image is the interference of rock joints and cracks

[0105] like Figure 6(a)-Figure 6(c) As shown in the figure, for the observation of the second type of image noise samples, it is found that the cracks and shapes on the rock surface are very similar to the blast hole traces, which leads to a low accuracy rate of feature extraction results of the ResUNet network and also misleads the manual labeling. Figure 6(c) shows that the detection results of the ResUNet-GhostNetV3 network of the embodiment of the present invention are affected by the fracture of the surrounding rock, and there is still an error with the standard result of the manual labeling Figure 6(b), but it is within the accuracy requirement of the construction.

[0106] Table 3 Model accuracy comparison data table / (m 3 )

[0107]

[0108] From the detection results of the above two types of noise images, it is found that the ResUNet network has certain limitations in extracting binary features of images, which causes certain interference to the detection accuracy of the GhostNetV3 network, but the detection accuracy remains at 82.3%, which can meet the requirements of construction accuracy. Due to the encoder-decoder and re-parameterized structure of the ResUNet-GhostNetV3 surrounding rock blast hole trace image detection network, the convolutional layer can automatically mine and learn the deep feature structure in the surrounding rock image, thereby reducing noise interference such as shadows and cracks, and has stronger generalization ability and robustness, which can significantly improve the detection accuracy.

[0109] The embodiment of the present invention realizes dynamic adjustment and optimization of drilling and blasting design by analyzing the measured cross-sectional data and based on the principle of similarity of surrounding rocks, provides an innovative solution for calculating the blast hole mark rate of surrounding rock in tunnels using the drilling and blasting method, provides construction personnel with a fast and accurate method for calculating the blast hole mark rate, and also lays a technical foundation for realizing comprehensive automation and refined blasting parameter control in the future.

[0110] In a specific embodiment of the present invention, a surrounding rock blasthole trace detection system based on the ResUNet-GhostNetV3 network includes:

[0111] A data set acquisition module, used to acquire surrounding rock image data sets;

[0112] A processing module, used for preprocessing the surrounding rock image data set, and using a ResUNet network to extract binary features from the surrounding rock image data set to construct a training data set;

[0113] Model building module, used to build a blasthole trace detection model based on GhostNetV3 convolutional neural network;

[0114] A training module, used for training and verifying the blast hole mark detection model through the training data set to obtain a final blast hole mark detection model;

[0115] The detection module is used to detect the blasthole marks in the surrounding rock through the final blasthole mark detection model.

[0116] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0117] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network, characterized in that: include: Acquire surrounding rock image dataset; Preprocessing the surrounding rock image dataset, and using a ResUNet network to extract binary features from the surrounding rock image dataset to construct a training dataset; Construct a blast hole trace detection model based on GhostNetV3 convolutional neural network; The blast hole mark detection model is trained and verified by using the training data set to obtain a final blast hole mark detection model; The blast hole traces in the surrounding rock are detected through the final blast hole trace detection model.

2. The surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network according to claim 1 is characterized in that: Preprocessing the surrounding rock image data set includes: The surrounding rock image dataset is annotated using an annotation tool to complete data annotation.

3. The surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network according to claim 1 is characterized in that: The ResUNet network includes an encoder and a decoder with jump connections; The encoder is used to extract abstract representations, and the path of the encoder includes: repeated convolution layers, batch normalization layers, and maximum pooling layers in sequence; The decoder is used to restore the original resolution, and the path of the decoder includes a transposed convolution layer and an upsampling layer.

4. The surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network according to claim 1 is characterized in that: Using a reparameterization method, linear parallel branches are introduced into depthwise convolution and 1x1 convolution when training the blast hole mark detection model.

5. The surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network according to claim 1 is characterized in that: The GhostNetV3 convolutional neural network includes multiple Ghost modules; The number of parameters and computational cost are reduced by the Ghost module.

6. The surrounding rock blasthole trace detection method based on the ResUNet-GhostNetV3 network according to claim 5 is characterized in that: The process of reducing the number of parameters and computational cost by the Ghost module includes: generating a cost feature Y, performing a deep convolution on the cost feature Y, and generating an original feature Y′; The original feature Y′ and the fixed feature are connected along the channel dimension to obtain the final output Y″, which is expressed as: Y = X * W; is the convolution kernel, X is the input feature, c in and c out Represent the input and output channel dimensions respectively, k is the kernel size, and * represents the convolution operation; Y′=X*W p ; Y″=Cat(Y′,X*W c ); in, Wp and W c They represent the main convolution operation and the parameters in the operation respectively, and Cat represents the connection operation.

7. The surrounding rock blast hole trace detection system based on ResUNet-GhostNetV3 network is characterized by: include: A data set acquisition module, used to acquire surrounding rock image data sets; A processing module, used for preprocessing the surrounding rock image data set, and using a ResUNet network to extract binary features from the surrounding rock image data set to construct a training data set; Model building module, used to build a blasthole trace detection model based on GhostNetV3 convolutional neural network; A training module, used for training and verifying the blast hole mark detection model through the training data set to obtain a final blast hole mark detection model; The detection module is used to detect the blasthole marks in the surrounding rock through the final blasthole mark detection model.

Images