Plastic Parts Surface Defect Detection Algorithm Based on Path-Augmentation Multi-Level Feature Pyramid Network

Abstract

Due to the characteristics of plastic parts surface defects such as variable scales and shapes, and a large number of small defect targets, existing algorithms cannot fully extract and utilize defect features. A plastic surface defect detection algorithm based on Path-Augmentation Multi-Level Feature Pyramid Network (PA-MLFPN) is proposed. On the basis of extracting multi-level and multi-scale defect features, to further enhance the representation ability of small defect targets, firstly, dilated convolution is adopted in the encoder of Thinned U-shape Modules (TUM) to fully retain shallow detail features. Meanwhile, a bottom-up feature augmentation path is added to the original TUM to interpolate shallow features into deep layers. Secondly, the Efficient Channel Attention (ECA) module is introduced in the second stage of the Scale-wise Feature Aggregation Module (SFAM) to achieve more reasonable weight allocation across channels. Finally, Focal Loss is used to calculate classification loss, which alleviates the problem of imbalance between positive and negative samples to a certain extent. Embedding the proposed PA-MLFPN into SSD and conducting experiments on the plastic parts surface defect dataset show that the algorithm achieves a mean Average Precision (mAP) of 84.89% and effectively solves the missing detection problem of small defect targets.