A fast and lightweight neural network for curve detection in structured-light vision sensing

Fringe signals like laser stripes exist in many fields, such as line structured light measurement, lane line detection, and so on. Line structured light measurement is an essential active vision technology, which is used in a variety of industry fields, such as chip appearance inspection and rail wear detection, in which the image of the light strip is very complex. Traditional methods fail to achieve high accuracy and high robustness when dealing with complicated environments. In this paper, a fast and robust method based on structured-light vision and deep learning is proposed. Our main contributions are: Firstly, we offer a fast laser stripe detection neural network (Fast-LSDNN). By applying a dual-channel network and knowledge distillation to our network, the speed and accuracy of laser stripe region positioning are greatly improved and carefully balanced. Secondly, we improve the traditional Gray-Gravity-method, and the running time is significantly reduced. We also build a dataset for chip appearance inspection independently, which contains different kinds of chip inspection scenes and can be transferred to other tasks directly by using our work. The experimental results show that our approach is robust and applicable.