Search

CN-121999363-A - Lightweight structure crack detection method and system based on cellular automaton

CN121999363ACN 121999363 ACN121999363 ACN 121999363ACN-121999363-A

Abstract

The invention belongs to the technical field of computer vision and structural safety monitoring, and particularly relates to a lightweight structural crack detection method and system based on cellular automata. The method comprises the steps of obtaining image data of a crack to be detected, preprocessing, inputting the preprocessed image into a backbone network, extracting a multi-scale feature map, compressing, evolving and expanding the feature map of each scale in the multi-scale feature map through cellular automaton evolution to obtain an enhanced multi-scale feature map, inputting the enhanced multi-scale feature map into a path aggregation network to conduct multi-scale feature fusion, inputting the fused feature map into a decoupling detection head, predicting the category confidence coefficient and boundary frame coordinates of the crack, and outputting the category and position information of the crack. The method and the device greatly reduce the parameter and the calculation complexity, and simultaneously remarkably improve the precision and the robustness of crack detection, and are suitable for real-time detection of edge equipment such as unmanned aerial vehicles.

Inventors

  • SHEN WENAI
  • ZHAO PENGHUI
  • GONG RUI
  • LONG ZHENTAO
  • LUO HUI

Assignees

  • 华中科技大学

Dates

Publication Date
20260508
Application Date
20260105

Claims (10)

  1. 1. The lightweight structural crack detection method based on cellular automaton is characterized by comprising the following steps of: step 1, acquiring crack image data to be detected and preprocessing; Step 2, inputting the preprocessed image into a backbone network, and extracting a multi-scale feature map; step 3, compressing, evolving and expanding the feature map of each scale in the multi-scale feature map through cellular automaton evolution to obtain an enhanced multi-scale feature map, wherein the evolution adopts a directional decoupling convolution evolution rule; Step 4, inputting the enhanced multi-scale feature map into a path aggregation network to perform multi-scale feature fusion; and 5, inputting the fused feature map into a decoupling detection head, respectively predicting the category confidence coefficient and the boundary frame coordinates of the crack, and outputting the category and the position information of the crack.
  2. 2. The cellular automaton-based lightweight structural crack detection method according to claim 1, wherein the backbone network adopts a pre-trained MobileNetV-Large network, and feature graphs with different downsampling multiplying power are respectively led out from different layers of the backbone network to obtain a multi-scale feature graph.
  3. 3. The cellular automaton-based lightweight structure crack detection method as claimed in claim 1, wherein in step 3, the cellular automaton evolution vs. input multiscale feature map Specifically, the compression, evolution and expansion of the feature channel comprises the steps of firstly compressing the feature channel through a dimension reduction unit: secondly, utilizing directivity decoupling convolution evolution rules in a low-dimensional space after dimension reduction And (3) carrying out T times of iteration evolution: And finally, restoring the evolved features to the original channel through the dimension raising unit and carrying out residual connection: In which, in the process, A 1 x 1 convolution is represented and, A batch normalization is shown and is performed, The activation function is represented as a function of the activation, Representing the directional decoupling convolution evolution rules, In order to provide an evolution step, The multi-scale characteristic map is output after cellular automaton evolution.
  4. 4. The cellular automaton-based lightweight structure crack detection method as set forth in claim 3, wherein the directional decoupling convolution evolution rule The method specifically comprises a vertical direction sensing branch and a horizontal direction sensing branch, wherein an input characteristic is marked as X, the convolution kernel size is k multiplied by k, and an evolution formula is as follows: in the formula, A vertical direction depth separable convolution representing a kernel size k x 1, A horizontal direction depth separable convolution representing a kernel size of 1 xk, For fusing the characteristic information of the two branches, And The output characteristics of the vertical and horizontal branches, respectively.
  5. 5. The cellular automaton-based lightweight structure crack detection method according to claim 3, wherein the iterative evolution adopts an asymmetric multi-step evolution strategy, and different evolution steps T are distributed for feature graphs of different scales; Sequentially marking the multi-scale feature map output by the backbone network as S3, S4 and S5 according to the spatial resolution from large to small, wherein the corresponding evolution steps are respectively as follows , , Then the following is satisfied: 。
  6. 6. The cellular automaton-based lightweight structure crack detection method as claimed in claim 3, wherein in the initial stage of model training, the convolution layer weights of the upgoing units in the evolution of the cellular automaton are weighted Bias term Initialized to zero, i.e.: 。
  7. 7. The cellular automaton-based lightweight structure crack detection method as in claim 5, wherein in step 4, the path aggregation network comprises a top-down upsampling fusion path and a bottom-up downsampling enhancement path; The top-down upsampling fusion path adds deep features and shallow features element by element in a channel dimension through upsampling; and the bottom-up downsampling enhancement path fuses shallow features and deep features through downsampling convolution operation, and finally, the feature graphs of the output unified channel numbers are respectively marked as N3, N4 and N5.
  8. 8. The cellular automaton-based lightweight structure crack detection method according to claim 7, wherein in step 5, the decoupling detection head respectively constructs a classification branch and a regression branch for the feature maps N3, N4, N5 of each level after the fusion; The classification branch comprises a plurality of 3 multiplied by 3 convolution layers and a1 multiplied by 1 convolution layer, and is used for outputting the confidence of the existence of the target; The regression branch comprises a plurality of 3×3 convolution layers and a1×1 convolution layer, and is used for outputting regression parameters of the target bounding box.
  9. 9. Lightweight structural crack detection system based on cellular automaton, characterized by comprising: The image acquisition module is used for acquiring the image data of the crack to be detected and preprocessing the image data; the backbone network module is used for carrying out feature extraction on the preprocessed image to obtain a multi-scale feature map; The cellular automaton evolution module is used for respectively compressing, evolving and expanding the multi-scale feature map to obtain an enhanced multi-scale feature map; The feature fusion module is used for inputting the enhanced multi-scale feature map into the path aggregation network to perform multi-scale feature fusion; And the decoupling detection module is used for respectively predicting the category confidence coefficient and the boundary frame coordinates of the crack according to the fused feature map so as to output the category and the position information of the crack.
  10. 10. The cellular automaton-based lightweight structure crack detection system of claim 9, wherein the cellular automaton evolution module comprises: The dimension reduction unit is used for compressing the characteristic channels of the multi-scale characteristic images; The evolution unit is used for performing iterative evolution by utilizing a directional decoupling convolution evolution rule, and the directional decoupling convolution evolution rule comprises a vertical direction sensing branch and a horizontal direction sensing branch; And the expansion unit is used for recovering the evolved characteristics to the original channel and carrying out residual connection.

Description

Lightweight structure crack detection method and system based on cellular automaton Technical Field The invention belongs to the technical field of computer vision and structural safety monitoring, and particularly relates to a lightweight structural crack detection method and system based on cellular automata. Background Cracks are a central concern in the health monitoring (SHM) of traffic infrastructure structures, such as roads, bridges, tunnels, etc., whose early discovery and repair is critical to the prevention of structural failure. Traditional crack detection mainly relies on manual visual inspection, and the mode has the obvious defects of low efficiency, strong subjectivity, high omission factor, high operation risk in dangerous areas and the like, and is difficult to meet the real-time and precision requirements of modern large-scale infrastructure maintenance. In recent years, with the rapid development of artificial intelligence technology, deep learning algorithms represented by Convolutional Neural Networks (CNNs), in particular YOLO (You Only Look Once) series object detection models, have been widely used in automatic structural crack detection due to their end-to-end detection capabilities. In the field of general target detection, in order to solve the problems of difficult multi-scale target detection and conflict of classification regression tasks, researchers propose a series of improved structures. For example, the path aggregation network (PANet) enhances the positioning information transmission capability of the feature pyramid by adding a bottom-up path, and the decoupling detection head (Decoupled Head) effectively solves the different requirements of the two tasks on the feature translation sensitivity by separating the classification task and the regression task into different branches. These advanced components have achieved significant success in generic object detection. However, in practical engineering applications, especially when real-time detection is performed by using an edge computing terminal such as an unmanned plane, a mobile detection vehicle or a handheld device, the prior art still faces serious challenges. The method mainly comprises the following steps of 1, enabling an existing high-performance detection model to have huge parameter quantity and high calculation complexity, and being difficult to realize real-time operation on edge equipment with limited resources and sensitive power consumption, compressing the model by greatly reducing the number of network layers or channels in order to adapt to a lightweight model designed by the edge equipment, so that the characteristic characterization capability of the model is obviously weakened, and being extremely easy to generate missed detection and false detection when facing to cracks under a fine or complex background, and 2, enabling a general convolution kernel to be mismatched with morphological characteristics of the cracks, which is a key technical bottleneck for limiting the precision of the lightweight model. Existing lightweight models mostly use standard 3 x 3 convolution kernels or depth separable convolutions. These convolution operators have the characteristic of being "isotropic" when aggregating spatial information, i.e., treating all directions within the receptive field equally. However, cracking is a very specific target for morphology, typically exhibiting long, continuous and highly directional linear characteristics. The general square convolution kernel is used for extracting the linear characteristic, so that the characteristic capturing efficiency is low, a large amount of irrelevant background texture noise is inevitably introduced, and the positioning accuracy of the model to the crack boundary is reduced. Therefore, there is an urgent need in the art for a new lightweight detection model that can break through the above bottleneck. Disclosure of Invention The invention aims to provide a lightweight structural crack detection method and system based on cellular automaton, which are used for carrying out operator-level optimization design aiming at anisotropic morphological characteristics of cracks, and remarkably improving sensitivity and detection precision of linear characteristics while maintaining extremely low parameter quantity and calculation quantity. In order to achieve the above purpose, the invention provides a lightweight structural crack detection method based on cellular automata, which comprises the following steps: step 1, acquiring crack image data to be detected and preprocessing; Step 2, inputting the preprocessed image into a backbone network, and extracting a multi-scale feature map; step 3, compressing, evolving and expanding the feature map of each scale in the multi-scale feature map through cellular automaton evolution to obtain an enhanced multi-scale feature map, wherein the evolution adopts a directional decoupling convolution evolution rule; Step 4, input