A novel defect detection system for complex freeform surface structures

Abstract In recent years, freeform surfaces have been widely used in various industries with the rapid development of manufacturing technology. However, the complex 3D structures and other surface properties of the freeform surface products challenge current visual inspection systems. To deal with the problem and meet the demand, we propose a novel, deep learning-based defect detection method for complex freeform surface structures in the injection industry. We first design a system of Multi-Lights Multi-Cameras with Polarizer, in which polarization is used to eliminate glare caused by freefrom surfaces and multi-illumination is used to enhance the visibility of certain defects. Then a convolutional neural network model based on YOLOv5 is utilized to detect small defects in images. Particularly, we raise several effective and easily-implemented strategies and improve the performance of the model significantly. Experimental results on our self-made dataset show that the improved model achieves a mean average precision (mAP) of 86.2% with 5.1% higher than the Baseline. Moreover, testing on the production line shows our detection system can benefit the related factory. Our work provides a promising approach to facilitate the development of visual inspection for products with complex freeform surfaces.