A multi-scale re-parameterization enhanced bilateral lightweight crack detection model for low-quality environments

Jingling Yuan, Nana Wang, Siqi Cai, Chunpeng Jiang, Xinping Li

Abstract: The detection of cracks in structural facilities is of significant importance for infrastructure maintenance and public safety. However, recent huge computations of deep neural networks make it difficult to run them on real-time monitoring devices with limited computing and storage capacity. Additionally, in many low-quality environments such as dim, shadows, and blurriness, lightweight models often lack the performance required for accurate detection of cracks. In view of the above, we propose the structural re-parameterization enhanced lightweight segmentation model for low-quality environments crack detection. The model adopts a bilateral structure and uses different multi-scale network construction units for the two branches of model, allowing for better crack feature extraction from different perspective to enhance robustness and performance in low-quality environments. We design specific re-parameterization process for each branch’s multi-scale construction unit to reduce the computational complexity of model inference, enabling the model to be better deployed on low resource monitoring devices for more efficient crack detection. A feature soft-selection mechanism is also proposed to better fuse the features of two branches. Experiment results on three concrete crack datasets indicate that our model can achieve good performance with lower complexity. Compared with the lightweight segmentation model BiSeNetV2, this model has nearly 52% and 43% less GFLOPs and MParams on the three datasets and achieves better performance.