Automated subway tunnel lining crack classification and detection based on two-step sequential convolutional neural network

Crack identification of subway tunnel lining is integral for preventive maintenance. The industry urgently requires accurate and efficient automation methods to handle the increasing length of subway lines. However, image-based crack identification accuracy in tunnel scenes is limited by low image contrast, uneven illumination, serious debris interference, poor connectivity integrity and other problems. Besides, when accuracy is pursued, models tend to become bloated and inefficient. This paper presents a two-stage sequential deep learning approach to solve the above issues. First, the integrated intelligent detection equipment is developed to realize the acquisition of high-definition image data. Second, a lightweight image classification network, MobileNet_WaCiCrack, is proposed to screen the images containing cracks, and then a multiscale crack detection network, YOLO_SWCrack, is developed to locate the cracks. Image classification and crack detection datasets are constructed for model training and validation. Experimental results show that the proposed method achieves high accuracy and efficiency and is superior to classical algorithms. MobileNet_WaCiCrack enlarges the receptive field through the wide area context information, avoiding the semantic feature loss caused by excessive compression and cropping of the image. YOLO_SWCrack introduces rich data augmentation technique, adaptive anchors, multi-scale detection heads, spatial feature pyramid modules, crack information integration modules, and wise-Io to form a multi-scale network architecture to extract abstract features from images and identify cracks in complex tunnel environments. The two-step sequential convolutional neural network effectively realizes intelligent identification of tunnel lining cracks, providing a powerful tool to assist sophisticated maintenance decisions.