文献类型：期刊文献

英文题名：Classification of Camellia oleifera using a dual recognition strategy based on deep learning

作者：Meng, Zhichao[1] Du, Xiaoqiang[1,2,3] Yao, Xiaohua[4] He, Leiying[1,2] Lin, Lepeng[5]

第一作者：Meng, Zhichao

机构：[1] School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou, 310018, China; [2] Zhejiang Key Laboratory of Transplanting Equipment and Technology, Hangzhou, 310018, China; [3] Provincial Key Laboratory of Agricultural Intelligent Sensing and Robotics, Zhejiang Province, Hangzhou, 310018, China; [4] Research Institute of Subtropical Forestry, Chinese Academy of Forestry, Hangzhou, 311400, China; [5] School of Youjia Intelligent Manufacturing, Hangzhou Vocational & amp, Technical College, Hangzhou, 310018, China

年份：2024

外文期刊名：Multimedia Tools and Applications

收录：EI(收录号：20242316193268);Scopus(收录号：2-s2.0-105004062418)

语种：英文

外文关键词：Commerce - Convolution - Crime - Deep learning - Learning systems

摘要：With the rapid growth of the Camellia oleifera industry, counterfeit seedlings have surfaced in the market. Authenticating the genuineness of Camellia oleifera species and cultivars has emerged as a substantial challenge. Present methods for discerning fake Camellia oleifera cultivars depend on manual expertise, which is both time-consuming and labor-intensive. Current deep learning-based crop classification methods can solely be used to classify training objectives and are unable to distinguish between authentic and counterfeit objectives. In light of this situation, it is crucial to establish an efficient recognition strategy to accurately recognize Camellia oleifera species and cultivars and thereby eliminate the distribution of fake seedlings in the market. Firstly, the image datasets of the foreside and backside of Camellia oleifera leaves are constructed in this study. The classification is accomplished using the EfficientNet model, and a dual recognition strategy is employed to verify fake Camellia oleifera. The experiments demonstrate that EfficientNet B4 achieves the best performance among the EfficientNet models. Furthermore, to further enhance the model's performance, the EfficientNet B4 model incorporates the Mish activation function, Multi-class Focal Loss, and SENet V2. The Improved EfficientNet B4 achieves an average T-acc of 97.5%, which is 0.89% higher than the original EfficientNet B4 for the training target classification. Additionally, for the non-training target, the Improved EfficientNet B4 achieves an average T-acc of 93.9%, surpassing the original EfficientNet B4 by 2.22%. The improved model is then compared with 11 other convolutional neural networks to highlight its superior performance. ? The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.