文献类型：会议论文

英文题名：Mushroom Image Classification and Recognition Based on Improved Swin Transformer

作者：Zhao, Kexin[1,2,3]; Huo, Yukang[1,2,3]; Xue, Lin[4]; Yao, Mingyuan[1,2,3]; Tian, Qingbin[1,2,3]; Wang, Haihua[1,2,3,5]

第一作者：Zhao, Kexin

机构：[1] National Innovation Center for Digital Fishery, Beijing, 100083, China; [2] Ministry of Agriculture and Rural Affairs, Key Laboratory of Smart Farming Technologies for Aquatic Animal and Livestock, Beijing, 100083, China; [3] China Agricultural University, College of Information and Electrical Engineering, Beijing, 100083, China; [4] Beijing Union University, Smart City College, Beijing, 100101, China; [5] Beijing Engineering and Technology Research Center for Internet of Things in Agriculture, Beijing, 100083, China

第一机构：National Innovation Center for Digital Fishery, Beijing, 100083, China

会议论文集：2023 IEEE 6th International Conference on Information Systems and Computer Aided Education, ICISCAE 2023

会议日期：September 23, 2023 - September 25, 2023

会议地点：Dalian, China

语种：英文

外文关键词：Classification and Recognition; Convolutional Neural Network; Deep Learning; Mushroom Image; Swin Transformer

摘要：Classification of mushrooms are essential for preventing even life-threatening consequences of accidentally eating wild mushrooms. In this study, a dataset including 114 varieties of mushrooms is collected and built. Next, the Swin Transformer model with robust classification characteristics is improved to classif mushroom images. Also, the classification accuracy of the improved Swin Transformer with different parameters was compared. In addition, ResNet50 is compared with the improved Swin Transformer under the optimal parameters. The model's training speed is further enhanced by integrating the improved Swin Transformer with ResNet50, thereby proposing the Swin_ResNet classification algorithm. Experimental results show that the classification accuracy of the optimal improved Swin Transformer algorithm is 87.66%, which is 14.86% and 7.64% higher than the ResNet50 and Swin Transformer models, respectively. In addition, the classification accuracy of Swin_ResNet is 85.31%, and the training time is 86.57% shorter than that of the improved Swin Transformer. This greatly improves the training efficiency. ? 2023 IEEE.