Lightweight computer vision based asphalt pavement crack visualization detection

Abstract

Cracks are the most common damage on urban asphalt roads, and it is important to detect cracks, which can effectively improve the safety performance of asphalt roads to extend their service life. The current detection of cracks mostly uses deep learning methods based on computer vision, and deep learning algorithms generally have the incompatibility problem of high accuracy and poor speed or fast speed and poor accuracy, and crack detection is limited by the demand of equipment can not be easily applied in the mobile terminal. Therefore, this paper proposes a lightweight road detection method based on computer vision. The method uses YOLOv5 as Baseline, and experiments are conducted by two different lightweight methods to explore the impact of pruning-based lightweight methods on network model size and detection speed and accuracy. The first method is based on weight pruning of BN layer, and the second method is based on pruning of Backbone's network structure. By comparing the evaluation with Baseline on the homemade dataset, the results show that the Backbone based network structure pruning only decreases the accuracy by 7.3% under the condition of pruning the Params by 55.2%, and the method achieves better results, which is beneficial to the fast and easy deployment of computer vision based deep learning networks in mobile.