YOLOv5-Sewer: Lightweight Sewer Defect Detection Model

Abstract

In the field of defect detection in sewers, some researches focus on high accuracy. However, it is challenging for portable on-site devices to provide high performance. This paper proposes a lightweight sewer defect detection model, You Only Look Once (YOLO) v5-Sewer. Firstly, the backbone network of YOLOv5s is replaced with a stacked MobileNetV3 block. Secondly, the C3 module of the neck of YOLOv5s is improved with a C3-Faster module. Thirdly, to compensate for the accuracy loss due to the lightweight network, a channel attention (CA) and convolutional block attention module (CBAM) are added to the proposed method. Finally, the Efficient Intersection over Union (EIOU) is adopted as the localization loss function. Experimental validation on the dataset shows that YOLOv5-Sewer achieves a 1.5% reduction in mean Average Precision (mAP) while reducing floating-point operations by 68%, the number of parameters by 55%, and the model size by 54%, compared to the YOLOv5s model. The detection speed reaches 112 frames per second (FPS) with the GPU (RTX 3070Ti). This model successfully implements a lightweight design while maintaining the detection accuracy, enhancing its functionality on low-performance devices.