Semantic segmentation method of hydraulic structure crack based on feature enhancement

Scientific, comprehensive, and standardized health monitoring is critical in the operation and maintenance of all types of water conservancy infrastructure. In this study, intelligent equipment is used to capture crack images of concrete dams and corridor hydraulic engineering scenes, and an artificial intelligence algorithm is used to achieve accurate recognition of crack information. However, most current research on concrete crack recognition lacks the analysis of crack information and simply obtains crack features through convolution and pooling to form a feature extraction network. The extracted high-dimensional features are not enhanced further, so the recognition effect cannot be continuously improved. A semantic segmentation technique for feature enhancement is proposed to solve the problem of low accuracy of crack location in the automatic detection of concrete cracks.

Statistical theory is used in this study to assess the pixel values of the cracked and non-cracked regions in three color channels and the proportion of the cracked region in the image. The size relationship and corresponding distribution of cracked and non-cracked regions on the pixel level are also obtained. Then, the ResNet-152 feature extraction network based on the residual structure is used to extract high-dimensional abstract semantic features from crack images. Due to the particularity of the residual structure, it can effectively reduce the loss of crack information during feature transmission and improve feature interoperability between different layers of the network so as to avoid the problem of gradient disappearance or explosion. Then, based on the results of statistical analysis, high-dimensional abstract features are sampled into two coarse segmentation feature maps corresponding to cracks and non-cracks. The similarity between the high-dimensional abstract features and the coarse segmentation feature map is calculated, the results of which are then used as weights to update high-dimensional abstract features to realize regional clustering of them. Finally, the clustered features are combined with the high-dimensional abstract features to obtain the enhanced features, which improve the crack location performance of the model. Meanwhile, the network loss function is optimized based on the crack information distribution. By controlling the number of samples used in the calculation of loss value, the contribution rate of crack information and non-crack information to the total loss value is balanced. As a result, the recognition accuracy of crack information is improved.

We used an unmanned aerial vehicle and an orbital robot to capture images of two hydraulic engineering scenes, including the dam and the corridor. After image preprocessing and labeling, we obtained a total of 3 000 crack images and labels, including 1 000 dam crack images and 500 corridor crack images. We stratified the data set into a training set, a validation set, and a test set in an 8∶1∶1 ratio. The crack pixel accuracy, recall rate, intersection-over-unions, and overall total pixel accuracy of the model on the test set reached 92.48%, 86.52%, 80.82%, and 99.79%, respectively.

By analyzing the relationship and distribution of pixel values between crack information and non-crack information in crack images and using them as prior information to construct a feature enhancement network and design the objective function of network optimization, the shortcomings of current concrete crack identification methods can be effectively overcome, and the performance of the network to recognize crack information can be improved.