外文摘要：There are more than 70 species of Vitis in the world, and 40 species, 1 subspecies and 13 varieties of Vitis native from China. Wild grapes are diverse and complex, and it is difficult to identify. As an important research direction, deep learning can quickly extract the deep features of images under the premise of ensuring high accuracy, and is widely used in the field of image classification and recognition. In this study, the images of mature leaves of 23 species of wild grapes under natural conditions were selected as the research objects, and the images of leaves were collected by smart phones to construct a wild grape leaf data set with a number of 10,077 images. Four deep learning networks (GoogleNet, ResNet-50, ResNet-101, VGG-16) were used to identify the leaf images of wild grapes. The classification network with the highest average accuracy is ResNet-101, average accuracy and recall rate are 98.64 % and 95.98 %, respectively. Among the 23 categories of the detection model, the prediction accuracy of 12 categories reached 100 %, and the recognition recall rate of 13 categories reached 100 %. It proves the feasibility of deep learning network model to identify leaf species in natural environment, and realizes automatic real-time identification of wild grapes, so as to provide reference for the protection, utilization, classification research of wild grapes and variety identification of other crops.
外文关键词：deep learning;species identification;Grad-CAM;Wild grapes
作者：Su, Baofeng;Sun, Lei;Fang, Yulin;Zhang, Ying;Pan, Bowen;Liu, Chonghuai;Ju, Yanlun;Fan, Xiucai;Jiang, Jianfu
作者单位：Chinese Acad Agr Sci;Northwest A&F Univ
期刊名称：SCIENTIA HORTICULTURAE
期刊影响因子：0.0
出版年份：2024
出版刊次：327
原文传递申请：江苏省科技资源（工程技术文献）统筹服务平台