基于ResNet和简化注意力机制的光盘识别算法  

Clean plate recognition algorithm based on ResNet and simplified attention mechanism

作　　者：李彦宏 薛佳炜 郭光远 林雨濛 王淑琴[1] LI Yanhong;XUE Jiawei;GUO Guangyuan;LIN Yumeng;WANG Shuqin(College of Computer and Information Engineering,Tianjin Normal University,Tianjin 300387,China)

机构地区：[1]天津师范大学计算机与信息工程学院,天津300387

出　　处：《天津师范大学学报（自然科学版）》2024年第2期69-74,共6页Journal of Tianjin Normal University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61070089);天津市应用基础与前沿技术研究计划重点资助项目(15JCYBJC4600);天津市科技计划资助项目(19JCZDJC35100).

摘　　要：提出一种基于ResNet的简化注意力机制的光盘识别算法.该简化注意力机制对结构相似的中间层共享同一注意力块,既保留网络对重要信息域的特征提取能力,又避免了注意力滥用造成的分类准确率下降和内存负载问题.使用光盘数据集(Plate Datasets)进行实验,结果表明,简化注意力的模型识别准确率相比一般注意力机制均有一定提升,并将注意力机制造成的额外参数量降低了72%.此外,在3个公开的十分类数据集上验证了简化注意力机制的有效性.A recognition algorithm for clean plate based on ResNet with simplified attention mechanism is proposed.The simplified attention mechanism shares the same attention block in the middle layer with similar structure,which not only pre-serves the feature extraction ability for important information domains,but also avoids the classification accuracy decrease and memory load caused by abuse of attention mechanism.The results of experiments on Plate Datasets show that,compared with the general attention mechanism,the recognition accuracy of models based on simplified attention is improved to some extent,and the number of additional parameters caused by attention mechanisms is reduced by 72%.In addition,the effectiveness of the simplified attention mechanism is verified on three public datasets with ten classifications.

关 键 词：图像分类 简化注意力机制 光盘识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]