摘要
针对民族药植物图像数据集稀缺、样本量少、图像背景复杂以致图像特征提取困难的问题,构建TibetanMP数据集,提出一种嵌入压缩—激励机制ResNet并结合迁移学习的图像识别方法。该方法对ResNet34在ImageNet上的预训练模型进行迁移学习,以减少过拟合现象,同时在网络浅层中引入SE机制,使网络聚焦图像中的关键特征,最后对模型进行微调。为了评估所提方法的性能,在TibetanMP、Oxford 102 flowers和CIFAR-10数据集上进行实验,模型分别取得96.33%、98.81%和91.92%的识别准确率。与其他主流CNN图像识别模型进行比较,发现该模型具有更高的识别精度,具有一定的工程实用性。
Aiming at the problems of scarcity of ethnic medicinal plants image data set, small sample size complex image background, which make image feature extraction difficult, the TibetanMP data set is constructed and an image recognition method with embedded Squeeze-andExcitation mechanism ResNet combined with transfer learning was proposed. In this method, the pre-training model of ResNet34 on ImageNet is transferred to reduce the over-fitting phenomenon. Meanwhile, SE mechanism is introduced in the shallow layer of the network to focus the key features in the image. Finally, the model is fine-tuned. In order to evaluate the performance of the proposed method, on the TibetanMP, Oxford 102 flowers and CIFAR-10 datasets, the model achieved recognition accuracy of 96.33%, 98.81% and 91.92%, respectively. Compared with other mainstream CNN image recognition models, this model has higher recognition accuracy. Experiments show that this method can effectively improve the image recognition performance of ethnic medicinal plants, and has certain engineering practicability.
作者
周婷
杜建强
朱彦陈
冯振乾
ZHOU Ting;DU Jian-qiang;ZHU Yan-chen;FENG Zhen-qian(College of Computer Science,Jiangxi University of Chinese Medicine,Nanchang 330004,China)
出处
《软件导刊》
2023年第2期1-7,共7页
Software Guide
基金
国家重点研发计划项目(2019YFC1712301)
国家自然科学基金项目(62141202)
江西省自然科学基金项目(20202BAB202019)
江西省一流学科建设科研启动基金专项(JXSYLXK-ZHYI060)
江西省教育厅科学技术研究项目(GJJ190683)
江西省教育厅科学技术研究项目(GJJ150862)
江西省研究生创新专项(YC2020-S370)。
关键词
图像识别
民族药植物图像
ResNet
压缩激励
迁移学习
image recognition
ethnic medicinal plants images
ResNet
squeeze and excitation
transfer learning
