CN-122023379-A - Engineering machinery crack detection method and system based on illumination decoupling and wavelet enhancement

CN122023379ACN 122023379 ACN122023379 ACN 122023379ACN-122023379-A

Abstract

The invention provides an engineering machinery crack detection method and system based on illumination decoupling and wavelet enhancement, and relates to the technical field of engineering machinery defect detection. The method comprises the steps of constructing a detection model, wherein the detection model comprises a feature coding network, a decoupling enhancement network and a feature decoding network, the feature coding network is used for extracting feature images with different scales, the decoupling enhancement network is used for decoupling illumination and reflection components of the feature images with different scales and enhancing wavelet frequency domain of the reflection components, the feature decoding network is used for generating a crack prediction mask by fusing the enhancement feature images with different scales, and the trained detection model is used for detecting a target image to be detected. The decoupling enhancement network adopts the Retinex decomposition module based on the adaptive adjustment of the scale parameters of the wavelet energy entropy, so that the problem of uneven illumination caused by strong reflection and complex shadow on the surface of engineering machinery is effectively solved, and the contrast ratio of micro crack components and the background is obviously improved while the interference of ambient light is inhibited.

Inventors

SHI HUAITAO
YANG HONGLEI
LI DONGJIE
ZHAO CHENGYING
CAI BING
GAO TIANHAO
YAN MING
ZHANG KE
BAI XU
WANG SHUYING

Assignees

沈阳建筑大学

Dates

Publication Date: 20260512
Application Date: 20260213

Claims (9)

1. The method for detecting the crack of the engineering machinery based on illumination decoupling and wavelet enhancement is characterized by comprising the following steps of: acquiring an original crack image, preprocessing the original crack image, and constructing a training set and a testing set based on the preprocessed crack image; The method comprises the steps of constructing a detection model, wherein the detection model comprises a feature coding network, a decoupling enhancement network and a feature decoding network, wherein the feature coding network is used for extracting feature graphs of different scales, inputting the feature graphs of different scales into the decoupling enhancement network, and the decoupling enhancement network is used for decoupling illumination and reflection components of the feature graphs of different scales and enhancing wavelet frequency domains of the reflection components to obtain enhancement feature graphs of different scales; Training and verifying the detection model by using the training set and the testing set to obtain a trained detection model; And detecting the target image to be detected based on the trained detection model to obtain a crack detection result.
2. The method for detecting the crack of the engineering machinery based on illumination decoupling and wavelet enhancement according to claim 1, wherein the feature coding network comprises a feature embedding layer and a plurality of cascaded structural interaction sensing modules, and the plurality of cascaded structural interaction sensing modules are used for extracting feature graphs with different scales; The feature embedding layer comprises a3 x 3 convolution for converting the preprocessed fracture image into an initial feature map of a specific dimension.
3. The method for detecting the cracks of the engineering machinery based on illumination decoupling and wavelet enhancement according to claim 2, wherein the structural interaction sensing module comprises a cavity space pyramid pooling submodule and a characteristic enhancement block; The cavity space pyramid pooling submodule comprises three parallel cavity convolution branches and a global average pooling branch, wherein each branch comprises at least one cavity convolution, different expansion rates are set for the three cavity convolution branches, the global average pooling branch comprises a global average pooling layer and an up-sampling layer, feature images extracted by the three cavity convolution branches and feature images output by the global average pooling branch are fused through channel splicing operation, and then the feature images are sequentially subjected to 1X 1 convolution dimension reduction, batch normalization layer and activation function processing to obtain an intermediate feature image output by the cavity space pyramid pooling submodule.
4. The method for detecting cracks of engineering machinery based on illumination decoupling and wavelet enhancement according to claim 3, wherein the feature enhancement block comprises a layer normalization, a 1 x 1 convolution layer, a 3x 3 depth separable convolution layer, a channel attention mechanism, a convolution layer and a Dropout layer, and the output of the Dropout layer and the input of the feature enhancement block are fused through residual connection; In the feature enhancement block, after the middle feature map output by the cavity space pyramid pooling submodule sequentially passes through a layer normalization layer, a1 multiplied by 1 convolution layer and a 3 multiplied by 3 depth separable convolution layer, the two sub feature maps are uniformly split into two sub feature maps along the channel dimension, the two sub feature maps are multiplied element by element, and sequentially pass through a channel attention mechanism, the convolution layer and a Dropout layer, and are fused with the middle feature map through residual connection, so that the feature map output by the feature enhancement block is obtained.
5. The method for detecting the cracks of the engineering machinery based on illumination decoupling and wavelet enhancement according to claim 1, wherein the decoupling enhancement network comprises a plurality of cascaded Retinex decomposition modules and a plurality of wavelet frequency enhancement modules, the Retinex decomposition modules at different levels respectively carry out smoothing processing on characteristic graphs output by a plurality of structural interaction perception modules in a characteristic coding network and extract reflection components, the output reflection components are input into corresponding wavelet frequency enhancement modules, and the wavelet energy entropy output by the wavelet frequency enhancement modules at different levels is input into a characteristic decoding network and a Retinex decomposition module at the next level; the Retinex decomposition module comprises a plurality of branches with different scales, each branch extracts a reflection component respectively, and the reflection components of the branches are spliced to obtain an enhanced reflection component output by the Retinex decomposition module; the specific method for extracting the reflection component from each branch is as follows: In each branch of the Retinex decomposition module, gaussian convolution operations with different scales are adopted for the feature images input into the Retinex decomposition module, so that illumination estimation components with corresponding scales are obtained; the characteristic diagram and the illumination estimation component are respectively compressed through a mapping function, and the compressed characteristic diagram and the illumination estimation component are subjected to differential operation to obtain a reflection component; The adaptive scale parameters of the one-dimensional filter kernel are adaptively adjusted through wavelet energy entropy output by the wavelet frequency enhancement module of the previous level.
6. The method for detecting the crack of the engineering machinery based on illumination decoupling and wavelet enhancement according to claim 5, wherein the wavelet frequency enhancement module comprises a discrete wavelet transform unit, a low-frequency enhancement unit, a high-frequency enhancement unit and an inverse discrete wavelet transform unit; The discrete wavelet transformation unit is used for decoupling the enhanced reflection component into frequency sub-bands with different scales by utilizing a wavelet basis function and calculating wavelet energy entropy, wherein the frequency sub-bands with different scales comprise a low frequency sub-band and a high frequency sub-band group; the high-frequency enhancement unit is used for sharpening the high-frequency sub-band and extracting a high-frequency characteristic diagram; The low-frequency enhancement unit is used for processing the low-frequency sub-band and extracting a low-frequency characteristic diagram; the inverse discrete wavelet transformation unit is used for reconstructing the high-frequency characteristic diagram and the low-frequency characteristic diagram back to a space domain to obtain an enhanced characteristic diagram.
7. The illumination decoupling and wavelet enhancement based engineering machinery crack detection method according to claim 6, wherein the feature decoding network comprises a plurality of cascaded hierarchical feature fusion modules and fusion convolution units; the hierarchical feature fusion module is used for carrying out deep semantic extraction and dimension recovery on the enhanced feature map; The fusion convolution unit is used for aggregating the outputs of the plurality of hierarchical feature fusion modules into a final prediction result; The multiple cascaded hierarchical feature fusion modules correspond to enhancement feature graphs with different scales, adopt a bottom-up cross-level fusion strategy, and integrate deep semantic information and shallow edge details among the enhancement feature graphs with different scales.
8. The method for detecting the crack of the engineering machinery based on illumination decoupling and wavelet enhancement according to claim 1, wherein the specific method for training and verifying the detection model by using the training set and the testing set is as follows: Comparing a pixel-level prediction probability map output by a feature decoding network with a label image corresponding to an original image, calculating a joint loss value based on a joint loss function, and judging whether the joint loss value meets a preset convergence condition; If the joint loss value does not meet the preset convergence condition, iteratively updating network parameters in the detection model through an error back propagation algorithm based on the joint loss value until the joint loss value meets the preset convergence condition; The preset convergence condition is that the parameter updating times reach the preset times or the joint loss value is smaller than a preset threshold value.
9. An engineering machinery crack detection system based on illumination decoupling and wavelet enhancement carries out engineering machinery crack detection based on the method of claim 1, and is characterized by comprising a crack image acquisition module, a structure interaction sensing module, a Retinex decomposition module, a wavelet frequency enhancement module, a hierarchical feature fusion module and a parameter optimization module; The crack image acquisition module is used for acquiring a surface original crack image of a key structure of the engineering machinery through unmanned aerial vehicle shooting or based on a structure detection report, wherein the original crack image comprises various crack morphology images under normal illumination, shadow shielding and weak light environment, and the original crack image is sent to the structure interaction sensing module; The structure interaction sensing module is used for extracting feature images with different scales of an original crack image to be detected; the Retinex decomposition module is used for decoupling illumination and reflection components of the feature images with different scales based on the adaptive scale parameters, and sending the decoupled reflection components to the wavelet frequency enhancement module; the wavelet frequency enhancement module is used for carrying out frequency domain decomposition and enhancement processing on the reflection component output by the Retinex decomposition module, obtaining an enhancement feature map and sending the enhancement feature map to the hierarchical feature fusion module; The hierarchical feature fusion module is used for conducting cross-level fusion and dimension recovery from bottom to top on the enhanced feature images with different scales to obtain a predicted feature image consistent with the surface original crack image scale, outputting a crack segmentation result of a pixel level, and sending the crack segmentation result to the parameter optimization module; The parameter optimization module is used for calculating a total loss value between a fracture segmentation result and a surface original fracture image, judging whether the total loss value meets a preset convergence condition, outputting a final fracture detection result under the condition that the preset convergence condition is met, and updating and iterating network parameters in the structure interaction sensing module, the Retinex decomposition module, the wavelet frequency enhancement module and the hierarchical feature fusion module according to the total loss value through a back propagation algorithm under the condition that the preset convergence condition is not met until the total loss value meets the preset convergence condition.

Description

Engineering machinery crack detection method and system based on illumination decoupling and wavelet enhancement Technical Field The invention belongs to the technical field of engineering machinery defect detection, and particularly relates to an engineering machinery crack detection method and system based on illumination decoupling and wavelet enhancement. Background The engineering machinery is used as core equipment for infrastructure construction, and the mechanical structure of the engineering machinery is under the action of heavy load and alternating stress for a long time, so that fatigue cracks are very easy to generate at key positions. It is counted that about 35% of engineering mechanical accidents are due to failure to find tiny cracks in the surface in time. These micro-cracks (typically no more than 0.3mm in width) can propagate rapidly under sustained stress, ultimately leading to catastrophic failure such as structural fracture. Therefore, the research on the surface crack detection method with high precision and strong robustness has important practical significance for guaranteeing the safe operation of mechanical equipment and preventing major accidents. The existing surface crack detection mainly comprises an artificial visual inspection method and an automatic detection method based on computer vision. Traditional manual visual inspection is greatly influenced by experience of detection personnel and has low efficiency, and objective evaluation is difficult to realize on a severe construction site. With the development of deep learning technology, the detection method based on semantic segmentation has significantly progressed in concrete structure and pavement crack detection due to the strong feature learning capability. However, most engineering machinery is made of metal, surface cracks of the engineering machinery are usually far finer than concrete cracks, the surface cracks are distributed in complex texture areas such as welding seam lap joints and corrosion points, and accurate positioning of micro features is difficult to achieve by a conventional detection model. In an actual engineering scenario, complex lighting conditions further limit the performance of existing detection models. On one hand, the strong reflection phenomenon of the metal surface often causes non-uniform illumination and local over-brightness of the image, so that edge details and contrast of micro cracks are weakened, and the traditional enhancing method based on the Retinex theory is difficult to balance between enhancing details and inhibiting reflection noise due to fixed parameters. On the other hand, the existing frequency enhancement method based on wavelet transformation is used for carrying out unified processing on each sub-band, and the energy distribution difference of crack characteristics in different frequency bands and directions is not fully considered, so that the problems of information redundancy or structural fracture are easy to occur under the condition of complex background interference. Therefore, how to realize effective decoupling enhancement of illumination interference and crack texture is a key challenge for improving the precision of micro crack detection. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides an engineering machinery crack detection method based on illumination decoupling and wavelet enhancement, which comprises the following steps: acquiring an original crack image, preprocessing the original crack image, and constructing a training set and a testing set based on the preprocessed crack image; The method comprises the steps of constructing a detection model, wherein the detection model comprises a feature coding network, a decoupling enhancement network and a feature decoding network, wherein the feature coding network is used for extracting feature graphs of different scales, inputting the feature graphs of different scales into the decoupling enhancement network, and the decoupling enhancement network is used for decoupling illumination and reflection components of the feature graphs of different scales and enhancing wavelet frequency domains of the reflection components to obtain enhancement feature graphs of different scales; Training and verifying the detection model by using the training set and the testing set to obtain a trained detection model; And detecting the target image to be detected based on the trained detection model to obtain a crack detection result. Further, the feature coding network comprises a feature embedding layer and a plurality of cascaded structural interaction sensing modules, wherein the plurality of cascaded structural interaction sensing modules are used for extracting feature graphs with different scales; The feature embedding layer comprises a3 x 3 convolution for converting the preprocessed fracture image into an initial feature map of a specific dimension. Further, the struct