Pavement Diseases Detection Using Improved YOLOv5

Abstract

Pavement diseases have a negative impact on traffic safety and ride comfort. With the rapid development of autonomous vehicle, the demand for rapid and accurate detection of pavement diseases is becoming more and more urgent. Previous pavement detectors have the contradiction between accuracy and speed. To address the above issue, a pavement disease detection model based on YOLOv5 is proposed. To improve the detection accuracy, we combine SPPF with attention mechanism, decouple the YOLOv5 detection head and use depthwise separable convolution. By using K-means to adjust the anchors, the convergence process of the model is smoother. The strategy of label smoothing is used to improve the generalization ability. Experiments on RDD2020 data set show that our method improves the accuracy of pavement diseases detection compared with the original YOLOv5 under the premise of maintaining real-time performance. Also the detection performance is better than EfficientDet, Faster RCNN and other series.