Ferrite Beads Surface Defect Detection Based on Spatial Attention Under Weakly Supervised Learning

Abstract

Ferrite beads’ automatic surface inspection is an important means to improve the quality and ensure proper operation. Although the deep-learning-based defect inspection methods reveal powerful performance, these methods often require a large amount of expensive annotation data for training, which limits the practical application of deep-learning-based inspection methods. To solve this problem, we propose a weakly supervised learning defect detection algorithm to achieve both high accuracy identification and effective localization of defects while using only image-level labels. To better use the texture location information of defects, we present a spatial association module (SAM) based on shallow features to improve the network performance. Then a training enhancement method is proposed to improve the detection ability, in which the guide crop and object ignore algorithms are used to extract the main defect area and background area in the image, respectively, to assist in generating optimal decisions. Finally, we put forward an optimal inference method to improve the completeness of localization without sacrificing accuracy, so as to provide a more reasonable and effective visual explanation for defect recognition. On the ferrite bead dataset, the proposed method uses less than 200 defect samples with only image-level labels in the training process to achieve a classification average precision (AP) of 97.1% with good stability and reliability, which has been used in the ferrite-bead inspection machine. To further verify the superiority and generalization, the proposed method is evaluated on several datasets for industrial quality inspection: Deutsche Arbeitsgemeinschaft fuer Mustererkennung (DAGM), KolektorSDD, and KolektorSDD2 achieve the best AP of defect classification of 100%, 100%, and 99.9%, respectively.