弱监督学习下的融合注意力机制的表面缺陷检测  

Weakly Supervised Surface Defect Detection Based on Attention Mechanism

作　　者：孙美君[1] 吕超章 韩亚洪[1] 李森[1] 王征[1] Sun Meijun;Lyu Chaozhang;Han Yahong;Li Sen;Wang Zheng(College of Intelligence and Computing,Tianjin University,Tianjin 300350)

机构地区：[1]天津大学智能与计算学部,天津300350

出　　处：《计算机辅助设计与图形学学报》2021年第6期920-928,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61772360,61876125).

摘　　要：现有基于深度学习的缺陷检测方法通常采用强监督学习策略,检测效果依赖于样本的数量和标注的质量.针对上述问题,提出弱监督学习下融合注意力机制的神经网络算法,仅使用图像级别标签便可同时预测缺陷的位置和概率.首先对多尺度感受野模块提取的特征应用特征融合网络,获取更多边缘细节信息;然后通过多层次的自编码器挖掘特征的深层语义信息;同时通过三线性全局注意力模块进一步细化浅层特征的空间位置信息;最后对浅层边缘特征和深层语义特征进行融合增强,得到最终的精细缺陷特征,达到高效准确的自动化表面缺陷检测的目的.基于PyTorch框架用KolektorSDD电转向器表面缺陷数据集验证所提算法,并与U-Net等缺陷检测算法进行对比.检测视觉效果显示,所提算法可以保留更多的细节纹理信息,能够有效扩大细微缺陷与复杂背景之间的特征差异.通过大量实验表明,该算法在复杂场景下比其他模型更为准确,其精准率、F1值和总体精度都有所提升.Nowadays,many defect detection algorithms based on deep learning are trained using a supervised learning strategy,which largely depends on the number of required samples and the quality of annotations.A weakly supervised attention network for surface defect detection is proposed,named SDD-Net,which can simultaneously predict both the location and probability of defects only by using image-level labels.Firstly,the feature fusion module is used to extract multi-scale edges features from the output of the multi-scale re-ceptive field.Then,the deep semantic information of features is mined by multilevel auto-coder.Meanwhile,the trilinear global context attention module is used to further refine the spatial location information of shallow features.Finally,the SDD-Net is used to integrate the shallow edge features and deep semantic fea-tures to obtain the final fine defect features.The results of evaluation on KolektorSDD dataset demonstrate that the SDD-Net based on the PyTorch framework has better detection performance than other detection methods such as U-Net.It can retain more detailed texture information and effectively expand the feature difference between small defects and complex background.The experimental results show that the proposed model is more accurate than the other models in the complex scene,the precision,the F1-score and the clas-sification accuracy are significantly improved.

关 键 词：深度学习 表面缺陷检测 弱监督学习 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]