A Lightweight Method for Detecting Sewer Defects Based on Improved YOLOv5

Abstract

In response to the issues of the existing sewer defect detection models, which are not applicable to small computing platforms due to their complex structure and large computational volume, as well as the low detection accuracy, a lightweight detection model based on YOLOv5, named YOLOv5-GBC, is proposed. Firstly, to address the computational redundancy problem of the traditional convolutional approach, GhostNet, which is composed of Ghost modules, is used to replace the original backbone network. Secondly, aiming at the problem of low detection accuracy of small defects, more detailed spatial information is introduced by fusing shallow features in the neck network, and weighted feature fusion is used to improve the feature fusion efficiency. Finally, to improve the sensitivity of the model to key feature information, the coordinate attention mechanism is introduced into the Ghost module and replaced the traditional convolution approach in the neck network. Experimental results show that compared with the YOLOv5 model, the model size and floating point of operations (FLOPs) of YOLOv5-GBC are reduced by 74.01% and 74.78%, respectively; the mean average precision (MAP) and recall are improved by 0.88% and 1.51%, respectively; the detection speed is increased by 63.64%; and the model size and computational volume are significantly reduced under the premise of ensuring the detection accuracy, which can effectively meet the needs of sewer defect detection on small computing platforms.