基于迁移学习的中国蛇类识别研究  

Research on Chinese Snake Recognition Based on Transfer Learning

作　　者：周志斌 罗志聪[1,2] 张展榜 孙奇燕[3] ZHOU Zhibin;LUO Zhicong;ZHANG Zhanbang;SUN Qiyan(College of Mechanical and Electronic Engineering,Fujian Agriculture and Forestry University,Fuzhou,350002,China;Fujian Key Laboratory of Agricultural Information Sensoring Technology,College of Mechanical and Electrical Engineering,Fujian Agriculture and Forestry University,Fuzhou,350002,China;College of Computer and Information Sciences,Fujian Agriculture and Forestry University,Fuzhou,350002,China)

机构地区：[1]福建农林大学机电工程学院,福州350002 [2]福建省农业信息感知技术重点实验室,福建农林大学机电工程学院,福州350002 [3]福建农林大学计算机与信息学院,福州350002

出　　处：《野生动物学报》2022年第2期436-443,共8页CHINESE JOURNAL OF WILDLIFE

基　　金：国家自然科学基金项目(61972093);海峡博士后交流资助计划项目。

摘　　要：蛇在野外广泛分布,不同种类的蛇具有不同的特性,实现蛇的准确识别对保护生物多样性和促进全球健康具有重要意义。为提高传统神经网络模型在蛇类图像上的识别效果,以中国地区常见蛇种作为研究对象,包括金环蛇(Bungarus fasciatus)、银环蛇(B.multicinctus)、圆斑蝰(Daboia russelli siamensis)、尖吻蝮(Deinagkistrodon acutus)、竹叶青(Trimeresurus stejnegeri)和王锦蛇(Elaphe carinata)。针对蛇类识别任务的特点,提出改进方案:根据蛇类形态特点,采用合适的数据增强方式扩充数据集;对神经网络模型进行改进,以适应蛇类图像分类任务;训练中采用Adam优化算法,优化模型学习过程;采用迁移学习方式,利用ImageNet数据集上训练得到的权值对模型初始化,并进行微调训练提升模型对蛇类识别任务的高阶特征表达能力。结果表明,改进方案可以有效提升卷积神经网络模型在蛇类图像识别任务上的准确度,在以VGG19、ResNet50、ResNet101、MobileNetV2和Xception网络为基础的分类模型上,平均识别准确率达到96.08%,可为建立蛇的自动识别系统提供参考。Snakes are widely distributed in the wild.Different snake species have different characteristics.Accurate identification of snakes is of great significance to protect biodiversity and promote global health.In order to improve the recognition effect of traditional neural network model on snake images,the common snake species in China were studied,including Bungarus fasciatus,B.multicinctus,Daboia russelii siamensis,Deinagkistrodon acutus,Trimeresurus stejnegeri and Elaphe carinata.According to the characteristics of snake recognition task,this paper proposed the following improvement scheme:based upon the characteristics of snake image,appropriate data augmentation method is used to expand the dataset;improved neural network model for snake image classification;Adam optimization algorithm optimized model learning process;using transfer learning,the weights trained on ImageNet dataset are used to initialize the model,and fine-tuning training is carried out to improve the high-order feature expression ability of the model for snake recognition task.The results showed that the improved scheme proposed in this paper could effectively improve the accuracy of convolutional neural network model in snake image recognition task.Using VGG19,ResNet50,ResNet101,MobilNetV2 and Xception networks,the average recognition accuracy could reach to 96.08%,which would provide a reference for the establishment of snake automatic recognition system.

关 键 词：蛇类识别 迁移学习 数据增强 微调训练 

分 类 号：Q958.1[生物学—动物学] TP391.4[自动化与计算机技术—计算机应用技术]