Tunnel lining surface defect detection algorithm based on multi-scale features and structural re-parameterization

Abstract To address the issues of low accuracy and speed in tunnel lining surface defect detection, an algorithm is proposed in which multi-scale feature extraction and structural re-parameterization are integrated for tunnel lining surface defect detection. Firstly, a grouped multi-kernel convolution block is designed to construct the multi-scale feature extraction module, which enhances the backbone network’s ability to detect surface defect features of varying scales, by capturing various types of tunnel lining surface defect features at multiple scales. Furthermore, the dilated reparam block and the generalized efficient layer aggregation network are combined to construct a novel dilated reparam block efficient layer aggregation network, which effectively fuses tunnel lining surface defect features of different scales extracted by the backbone network, and through a structural re-parameterization strategy, equivalently transforms the multi-branch dilated convolution structure into a single large non-dilated convolutional kernel, reducing computational complexity and improving inference speed. Finally, an occlusion-aware detection head is designed to perform channel weighting on the fused features, by enhancing the model’s attention to important defect feature information, the problem of occlusion and overlap in tunnel lining surface defects is alleviated, thereby improving the accuracy of defect detection. Experimental results show that, compared to the baseline model, the number of parameters and the computational cost in the proposed algorithm are reduced by approximately 13.5% and 15.9% respectively, the F1-score is increased by 3.2%, and the mean average precision ([email protected]) for tunnel lining surface defect detection is improved by 2.6%, and the inference speed is increased by approximately 13 frames per second.