基于光度立体和深度学习的电池缺陷检测方法  

Deep photometric stereo learning framework for battery defect detection

作　　者：苏育挺 景梦瑶 井佩光 刘先燚 SU Yu-ting;JING Meng-yao;JING Pei-guang;LIU Xian-yi(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072

出　　处：《吉林大学学报(工学版)》2024年第12期3653-3659,共7页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金青年项目(61802277);天津市自然科学基金项目(20JCQNJC01210)。

摘　　要：针对电池缺陷检测易受黑色外观干扰,导致仅通过一张单光源下观测图像的局限视觉,无法实现缺陷的有效识别的问题,提出了一种端到端的光度立体视觉缺陷检测模型。首先,利用光度立体特征生成模块生成法线特征,获取物体表面细节信息;然后,采用通道协同注意力机制,探讨特征信道间的相互关系,充分挖掘特征间的关联以自适应地增强全局表示,进一步提升信息表达能力;最后,利用特征金字塔和空间金字塔池化实现多尺度预测,提升分类准确率。在自建Battery 101数据集上的实验结果表明:与其他算法相比,本文算法在检测精度和推理速度上都取得较好效果。此外,消融实验也进一步验证了模型中各个模块的有效性。Aiming at the problem that the battery defect detection is susceptible to black appearance interference,which leads to the limited detect identification of the observation image under a single light source,an end-to-end deep-learning photometric stereo network(DPSNet)is proposed.Firstly,the photometric stereo feature generator(PSFG)enables model to facilitate transformation from multi-input features to normal feature so that better excel in defect signals.Then channel co-attention(CCA)module explores the channel interactions between each input and surface normal for informative representations,and fuses modify features to coordinately enhance global representation.Finally,spatial pyramid pooling(SPP)and feature pyramid networks(FPN)achieve multi-scale prediction.The experimental results on the self-built Battery101 dataset show that the proposed method achieves better effect.In addition,the ablation experiment further verifies the effectiveness of each module in the model.

关 键 词：信号与信息处理 电池缺陷检测 光度立体视觉 通道协同注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]