Segmentation Method of Magnetic Tile Surface Defects Based on Deep Learning

Abstract

Magnet tile is an essential part of various industrial motors, and its performance significantly affects the use of the motor. Various defects such as blowholes, break, cracks, fray, uneven, etc., may appear on the surface of the magnet tile. At present, most of these defects rely on manual visual inspection. To solve the problems of slow speed and low accuracy of segmentation of different defects on the magnetic tile surface, in this paper, we propose a segmentation method of the weighted YOLACT model. The proposed model uses the resnet101 network as the backbone, obtains multi-scale features through the weighted feature pyramid network, and performs two parallel subtasks simultaneously: generating a set of prototype masks and predicting the mask coefficients of each target. In the prediction mask coefficient branch, the residual structure and weights are introduced. Then, masks are generated by linearly combining the prototypes and the mask coefficients to complete the final target segmentation. The experimental results show that the proposed method achieves 43.44/53.44 mask and box mAP on the magnetic tile surface defect dataset, and the segmentation speed reaches 24.40 fps, achieving good segmentation results.