A Stable Lightweight Model for Metal Crack Detection Based on YOLOv5

Abstract

With the development of the world’s industrialization, more and more metal parts are used as important supports or working parts in machines. Due to the extension of use time, metal parts gradually fatigue and develop cracks, requiring repairs. The current repairing methods are mainly through manual visual inspection, physical methods or object detection algorithm. However, manual visual inspection has the risk of omission, the use of physical methods is costly, low accurate, and shows poor real-time performance, and the current object detection algorithm cannot meet the need of both rapidity and accuracy at the same time. In order to solve the problem that the current algorithm cannot satisfy the requirement of real-time and accuracy at the same time, we design a stable lightweight model based on the YOLOv5. The model adds a quadruple down-sampling feature extractor, puts cross-layer connection lines between Head and Backbone, increases the SE attention mechanism, and adjusts the loss function. The experimental results show that this model has higher detection accuracy in metal crack detection on the basis of ensuring real-time performance.