One-stage object detection networks for inspecting the surface defects of magnetic tiles

One of the core components of the permanent magnet motor is magnetic tile and surface defect detection of it is of vital importance to ensure the performance and service life of the motor. This paper employs a deep learning method based on computer vision to detect the surface defects on magnetic tiles in order to replace manual inspection and increase productivity. Considering the real-time requirements of the industrial site, three designed one-stage object detection networks of different depth are compared on our Inner-R surface dataset of magnetic tiles. The whole image is input into the networks which regard the object detection as a regression problem and output the value of class probability and position coordinate of the object. This approach can detect more than one defects on the same image as well as the location of defects which provides advantages to find the number of defects per class and improve the manufacturing process. As the result shows, the YOLOv3 network is the most applicable one in this magnetic tile surface defect detection problem and the detection time is less than 23 ms, which is an eye-catching result.