Intelligent recognition system and detection algorithm optimization for urban tunnel diseases based on convolutional neural network

In order to solve the problems of insufficient recognition accuracy, low robustness, and slow detection speed in existing tunnel face joint and crack recognition methods, the author proposes an intelligent recognition system and detection algorithm optimization for urban tunnel diseases based on convolutional neural networks. The author uses the ResNet-18 residual neural network, which balances inference speed and prediction accuracy, as the backbone structure, and replaces standard convolution with depthwise separable convolution to build a lightweight model ResNet DS (Depthwise Separable) suitable for real-time screening. The model is optimized using weight loss function and static offline quantization. The new algorithm lightens the neural network and achieves high-speed recognition and removal of massive images. The experimental results show that the improved lightweight model has a recognition accuracy of up to 98.66%. Compared with the original model, the filtering speed is improved by 22% (10.86ms/sheet) on GPU (RTX20600Super8G) and 178% (21.20 ms/sheet) on CPU. This study provides a real-time screening algorithm for disease rapid collection systems, which better meets detection needs.