An Analysis of Defect Detection on Steel Strip Surface Using Modified Yolo V5

Abstract

Defects in any type of material do lessen the strength and will diminish the service life which in-turn affects the quality. Steel surface images with broad scale variance and complicated background lessens the high quality and productiveness. In steel surface images detection, there are hitches of small target sizes and imprecise features that causes incorrect and ignored detection. Nevertheless, these problems can be avoided with the aid of inspecting the defects earlier than processing with accurate and automated object detection algorithms. YOLOv5, a fine tuned learning model is evaluated on available steel strip datasets, which solves identification and detection problems in industrial scenario. The current YOLOv5 research focus on improving accuracy on defect classification and detection speed however there’s no distinct records on algorithm execution time and size. This study focus on YOLOv5 model with induced GHOST and TRANSFORMER modules. Through the new modules, we are looking to reduce the size of the algorithm, number of parameters and number of GFLOPS. By these changes this model may be deployed easily on mobile, embedded devices for detection of small, easily confused defects.