CN-121686253-B - Ground penetrating radar image disease detection method, system and electronic equipment

CN121686253BCN 121686253 BCN121686253 BCN 121686253BCN-121686253-B

Abstract

The invention discloses a method, a system and electronic equipment for detecting image diseases of a ground penetrating radar, and belongs to the technical field of computer vision and road disease detection. The method comprises the steps of obtaining and preprocessing a GPR image, constructing a disease detection model, wherein the model comprises a multi-scale vertical/horizontal stripe convolution module, a self-adaptive gating routing module and a bidirectional attention fusion module which are respectively used for extracting directional characteristics, dynamically fusing multi-scale characteristics and self-adaptive fusion direction characteristics, and training and utilizing the model to identify and position the disease. According to the method, the computational complexity is reduced through the stripe convolution, the self-adaptive feature fusion is realized through the gating and attention mechanism, and the detection precision and efficiency of multi-scale and directional disease targets in the GPR image are obviously improved.

Inventors

YU XIN
QIU HAOYU
LI DI
ZHANG JIN
CHEN RUIKANG
PENG YAOWEI
LIU WEIZHEN
FENG YUYANG

Assignees

长沙理工大学

Dates

Publication Date: 20260505
Application Date: 20260211

Claims (7)

1. The ground penetrating radar image disease detection method is characterized by comprising the following steps of: s1, acquiring ground penetrating radar image data and preprocessing; S2, constructing a disease detection model, wherein the disease detection model comprises a multi-scale vertical stripe convolution module, a multi-scale horizontal stripe convolution module, a self-adaptive gating routing module and a bidirectional attention fusion module; The multi-scale vertical stripe convolution module is used for extracting longitudinal features, and adopts at least three vertical direction one-dimensional stripe convolution kernels with different scales to extract the multi-scale longitudinal features in parallel, wherein the shape of each vertical direction one-dimensional stripe convolution kernel is (k, 1), and k is a scale parameter; The multi-scale horizontal stripe convolution module is used for extracting transverse characteristics, the multi-scale horizontal stripe convolution module adopts at least three horizontal one-dimensional stripe convolution kernels with different scales to extract the multi-scale transverse characteristics in parallel, the shape of each horizontal one-dimensional stripe convolution kernel is (1, k), and k is a scale parameter; The self-adaptive gating routing module is used for respectively carrying out dynamic weighted fusion on the multi-scale characteristics in the vertical direction and the multi-scale characteristics in the horizontal direction to generate fused longitudinal characteristics and fused transverse characteristics, and the self-adaptive gating routing module executes the following steps: Stacking output features of the multi-scale branches in the same direction as feature tensors; carrying out global average pooling on the stacked features to obtain channel level description vectors; generating weight fractions of all branches through two layers of 1×1 convolution; Normalizing the weight fraction by using a Softmax function with temperature parameters to obtain the fusion weight of each branch; carrying out weighted summation on the multi-scale features according to the fusion weights to obtain the fused direction features; The two-way attention fusion module is used for carrying out self-adaptive weighted fusion on the fused longitudinal characteristics and the fused transverse characteristics to generate final fusion characteristics, and the two-way attention fusion module executes the following steps: splicing the fused longitudinal features and the fused transverse features along the channel dimension; global average pooling is carried out on the spliced features, and two weight values are generated through a two-layer convolution network; Normalizing the weight value by using a Softmax function to obtain a fusion weight of the longitudinal feature and the transverse feature; weighting and fusing the longitudinal features and the transverse features according to the fusion weights to obtain the final fusion features; S3, training the disease detection model by using the preprocessed data; S4, performing disease identification and positioning on the ground penetrating radar image to be detected by using the trained model.
2. The method of claim 1, wherein the multi-scale vertical stripe convolution module and the multi-scale horizontal stripe convolution module are each implemented using a depth separable convolution, the convolution kernel scale of which comprises at least three of 3, 7, 15, 19, 31.
3. The method according to claim 1, wherein the temperature parameter is used to adjust the smoothness of the weight distribution.
4. The method of claim 1, wherein the impairment detection model is framed in a YOLOv m network structure and replaces C2f modules therein with bi-directional fusion modules comprising the multi-scale vertical stripe convolution module, the multi-scale horizontal stripe convolution module, the self-adaptive gating routing module, and the bi-directional attention fusion module.
5. A ground penetrating radar image disease detection system, comprising: the data acquisition and preprocessing unit is used for acquiring and preprocessing the image data of the ground penetrating radar; a model construction unit for constructing the disease detection model according to any one of claims 1 to 4; The training unit is used for training the disease detection model by using the preprocessed data; And the detection unit is used for identifying and positioning diseases of the image to be detected by using the trained model.
6. The system of claim 5, wherein the multi-scale vertical stripe convolution module and the multi-scale horizontal stripe convolution module in the lesion detection model have convolution kernel scales configured as adjustable parameters including at least three of 3, 7, 15, 19, 31.
7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, the processor implementing the ground penetrating radar image fault detection method of any one of claims 1 to 4 when executing the computer program.

Description

Ground penetrating radar image disease detection method, system and electronic equipment Technical Field The invention relates to the technical fields of computer vision, deep learning and nondestructive detection, in particular to a method, a system and electronic equipment for detecting image diseases of a ground penetrating radar. Background Ground Penetrating Radar (GPR) is a technology widely applied to nondestructive testing of infrastructure internal structures such as roads, bridges and the like. By interpreting the acquired GPR image, diseases such as cracks, interlayer defects, looseness and the like can be identified. However, automated lesion detection of GPR images presents challenges in that lesion targets typically exhibit significant directionality (e.g., longitudinal cracks, lateral delamination) and anisotropy, lesions of different depths and types vary greatly on the image scale, and interference such as electromagnetic noise, clutter, and multiple reflections that are prevalent in the image severely masks the lesion features. When the existing general target detection algorithm is directly applied to the GPR image, the problems of low detection precision, poor generalization capability, high calculation cost and difficulty in meeting engineering real-time requirements are caused by the fact that a calculation convolution kernel receptive field is fixed, the direction modeling capability is insufficient, the sensitivity to noise is high and the like. Disclosure of Invention The invention aims to overcome the defects of the prior art and provides a method, a system and electronic equipment for detecting the image disease of the ground penetrating radar, which have the advantages of high detection precision, excellent calculation efficiency and sensitivity to the directional disease. In order to achieve the above purpose, the invention adopts the following technical scheme: In a first aspect, the present invention provides a method for detecting an image disease of a ground penetrating radar, comprising the steps of: s1, acquiring ground penetrating radar image data and preprocessing; S2, constructing a disease detection model, wherein the disease detection model comprises a multi-scale vertical stripe convolution module, a multi-scale horizontal stripe convolution module, a self-adaptive gating routing module and a bidirectional attention fusion module, wherein the multi-scale vertical stripe convolution module is used for extracting longitudinal characteristics, the multi-scale horizontal stripe convolution module is used for extracting transverse characteristics, the self-adaptive gating routing module is used for carrying out dynamic weighted fusion on the multi-scale characteristics, and the bidirectional attention fusion module is used for carrying out self-adaptive weighted fusion on the characteristics in the vertical and horizontal directions; S3, training the disease detection model by using the preprocessed data; S4, performing disease identification and positioning on the ground penetrating radar image to be detected by using the trained model. Further, the multi-scale vertical stripe convolution module and the multi-scale horizontal stripe convolution module are both realized by adopting depth separable convolution, and the convolution kernel scale comprises at least three of 3, 7, 15, 19 and 31. The design can capture multi-scale characteristics from local details to a macroscopic structure at the same time with lower parameter quantity and calculation cost, and obviously improves the perception capability of the model on diseases with different sizes. The method comprises the steps of obtaining a channel level description vector by carrying out global average pooling on the characteristics after stacking, generating weight fractions of all branches through two layers of 1X 1 convolution, normalizing the weight fractions by using a Softmax function with temperature parameters to obtain fusion weights of all branches, and carrying out weighted summation on the multi-scale characteristics according to the fusion weights to obtain the fused characteristics. The mechanism enables the network to adaptively select and fuse the most relevant scale features according to the content of the input image, and enhances the feature selection capability and robustness of the model under the complex noise background. Further, the temperature parameter is used for adjusting the smoothness degree of the weight distribution, so that the degree of 'softness' of feature fusion is flexibly controlled. The bidirectional attention fusion module comprises the following steps of splicing vertical branch output features and horizontal branch output features along a channel dimension, carrying out global average pooling on the spliced features, generating two weight values through a two-layer convolution network, normalizing the weight values by using a Softmax function to obtain fusion weights of the ve