Automatic Surface Defect Inspection System Using Convolutional Neural Networks

Abstract

Abstract Surface quality of the piston rod is very important for the durability of the shock absorbers. However, currently in production, it is very difficult to use the available systems to inspect the whole rod surface online, in real time (∼4s), with a high sensitivity and accuracy. To overcome this, in this paper, an online automatic rod inspection system has been developed, which allows to inspect the whole rod surface in 4s, with a sizing accuracy around 75%, and detection accuracy <90%. An integrated software is used for rod rotation control, image capturing, image processing, and decision making. Convolutional Neural Network is used to processing the 360° surface image with a high accuracy, eliminating errors caused by environmental lighting. A new method based on aspect ratio and size information is used for defects classification. Experimental results show that the system is capable of detecting defects as small as 25μm and differentiating nodule, dent, and scratch with a processing time around 4s per rod.