Few-shot Learning Combine Attention Mechanism-Based Defect Detection in Bar Surface

Defect detection on bar surface is a challenging task due to the complex and variable bar surface conditions. Traditional pattern recognition methods are widely used to detect defects in the industry, however most of existing methods are not very universal for all kinds of defects. Meanwhile because...

Full description

Saved in:
Bibliographic Details
Published in:ISIJ International 2019/06/15, Vol.59(6), pp.1089-1097
Main Authors: Lv, Qianwen, Song, Yonghong
Format: Article
Language:English
Subjects:
attention mechanism
bars
few-shot learning
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Defect detection on bar surface is a challenging task due to the complex and variable bar surface conditions. Traditional pattern recognition methods are widely used to detect defects in the industry, however most of existing methods are not very universal for all kinds of defects. Meanwhile because of the limited number of defective samples, traditional deep learning methods are not very effective in practice. This paper addresses these issues and proposes a novel few-shot learning method which combines with attention mechanism. Our method is built by a Convolutional Neural Network (CNN) which extracts image features, and a Relation Network (RN) which calculates the similarity score between a pair of images, predicts image categories through similarity scores. Firstly, in order to extract more effective and discriminative features, we introduced Squeeze-and-Excitation Networks (SENet) as an attention module into our method which can enhance effective features and weaken invalid features. Secondly, unlike traditional object detection techniques which mainly focus on foreground information, background information is also necessary in our method, because we need to utilize background information to distinguish pseudo and real defects. So in our method, we replaced Max-Pooling with Mean-Pooling. Finally, in order to solve the low efficiency of parameter update caused by sharp dropping of loss function values on our dataset, we use L1Loss and BCELoss to replace Mean square error loss function. Experiment results show that the proposed method can achieve an average accuracy rate of 97.25% on our data set, increased by 7.92% compared with state-of-the-art.
ISSN:0915-1559
1347-5460
DOI:10.2355/isijinternational.ISIJINT-2018-722