Detection of weak micro-scratches by semantic segmentation

Abstract

A machine vision-based deep learning is designed for detecting scratches on aspherical lenses' surfaces. This system includes a mechanical module integrated with a hybrid lighting system, and an industrial camera with micro-lens to automatically examine the surface of a lens. The entire surface is collected, the scratches dataset is established and manual annotation is performed as a training dataset to feed into a convolutional neural network. A deep learning model is introduced based on the DeepLabv3 architecture to automatically detect and expose scratch locations by segmenting their shapes. The experimental results show that the model outperforms traditional computer vision algorithms and other state-of-the-art neural networks, achieving a segmentation accuracy of 86%. Besides, the machine vision system reaches an impressive detection speed of 0.9 s per image.