Multi-task deep learning-based surface crack detection for Molded Power Inductor

Abstract

Molded power inductor is a widely used electronic component, in its production process, there will be cracks, scratches, fragments and other surface defects, affecting product quality, so it is necessary to conduct surface defect detection. Among them, crack defects are very subtle and rare, which brings great challenges to detection. This paper proposes a multi-task deep learning based detection method, which effectively solves the problem of inductive crack defect detection in the case of small sample, which adopts U-Net network architecture, uses Resnet as the encoder backbone network, and connects different task-related networks at the decoding end, including crack area detection task, crack centerline detection task and crack binary classification task network, and improves the detection effect of crack through multi-task learning. The test results show that compared with the single-task method, the performance index of crack area detection in this method is improved by 3%, according to the presence or absence of cracks, the classification accuracy rate of good and defective products is 93.1%, and the recall rate of defective products is 100%, which meets the actual project detection needs.