Lightweight Road Pothole Detection Network Based on FCOS

To address the problem of low efficiency and high danger of manual pothole detection on roads in road maintenance work, we propose a lightweight pothole detection algorithm based on FCOS, which use convolutional neural networks to automatically detect potholes on the road, keeping the network lightweight while ensuring high accuracy. The main ideas of our algorithm are as follows: (1) We use edge segmentation to extract the texture information of potholes from the image as a priori knowledge, and then fuse this information with the original image and input it into the backbone network to enhance the feature extraction capability of the network for the region of interest. (2) The ResNet backbone network of the baseline FCOS is replaced by a G-GhostNet network structure, which greatly reduces the model training overhead and parameter settings. (3) We add the SCConv module after the backbone network to improve spatial and channel feature extraction and reduce feature redundancy. The results of the experiments verify that our algorithm outperforms baselink FCOS with 1.4 % improvement in mAP, 30.9% reduction in the number of model parameters and 27.5% reduction in computational overhead. By contrasted with some state-of-the-art models, the proposed network has higher accuracy, better generalization capability, less computational resources, and better device compatibility. The proposed algorithm contributes to the application of smart road pothole detection techniques on mobile and embedded devices.