Road Crack Detection System Using Image Segmentation Algorithm

Road surface cracks pose a significant potential threat to driving safety, and previous manual detection methods have low efficiency. The existing crack detection methods have low generalization ability, poor crack segmentation ability and efficiency in complex backgrounds. In response to the shortcomings of traditional road crack detection systems, this article combines the current image segmentation algorithms in the field of deep learning to design and implement an intelligent road crack detection system based on deep learning technology. The system solves the problem of difficult deployment of deep learning algorithms in power constrained terminals, and it optimizes the model and accelerates inference through the OpenVINO platform. Secondly, this article proposes a new improved network structure, Crack U-Net, based on an encoder decoder structure, with the aim of improving the model generalization and detection accuracy of road crack detection. Crack U-Net uses a Crack U-block composed of residual blocks and mini-U as the basic convolutional module of the network. Compared to traditional double-layer convolutional layers, Crack U-block can extract richer crack features. To verify the effectiveness of the Crack U-Net model, a series of tests were conducted on a publicly available crack dataset. The experimental results show that the AIU value of Crack U-Net on the dataset has increased by 2.5% compared to previous methods, and it is superior to existing methods in crack segmentation accuracy and generalization. In addition, the experimental results of the parameter reduction section have shown that Crack U-Net can perform significant model pruning, and mobile devices such as drones can meet the computational resources required for the pruned Crack U-Net model.