基于深度卷积神经网络与迁移学习的鱼类分类识别  

Fish Classification Based on Deep Convolutional Neural Network and Transfer Learning

作　　者：陈文辉 蔡妹姝 严松 张宏 徐琳[1] 林茂兹[1] CHEN Wenhui;CAI Meishu;YAN Song;ZHANG Hong;XU Lin;LIN Maozi(Fuqing Branch of Fujian Normal University Institute of Ocean Research,Fuqing,Fujian 350300,China;Fuqing Branch of Fujian Normal University Key Laboratory of Nondestructive Testing,Fuqing,Fujian 350300,China;Fuqing Branch of Fujian Normal University School of Ocean Science and Biocchemistry Engineering,Fuqing,Fujian 350300,China)

机构地区：[1]福建师范大学福清分校海洋研究院,福建福清350300 [2]福建师范大学福清分校无损检测技术福建省高校重点实验室,福建福清350300 [3]福建师范大学福清分校海洋与生化工程学院,福建福清350300

出　　处：《福建师大福清分校学报》2019年第5期30-38,共9页Journal of Fuqing Branch of Fujian Normal University

基　　金：2019年度福建省海洋经济发展补助资金项目(FJHJF-L-2019-7);中央引导地方科技发展专项(2019L3013);福建省中青年教师教育科研项目(JT180631,JT180632);福建省教育科学“十三五”规划课题(FJJKCGZ17-194);福建师范大学福清分校科研创新培育项目(科技类)(KY201707)

摘　　要：高效的鱼类分类识别是海洋牧场智能化监测的基础.传统的通过浅层模型,利用目标特征的分类识别方法效率低下,泛化性差,难以实现智能化应用;而重建并训练深度卷积神经网络(DCNN)模型占用巨大的计算机资源.文章提出一种基于DCNN和迁移学习的方法,针对新图像数据集,通过选择训练参数,对预训练模型进行再训练,实现鱼类的分类识别.通过实验证实,这种方法可在占用少量的计算机资源情况下,达到97.14%的验证准确率.使用基于DCNN与参数迁移的学习策略可以得到性能良好的深度神经网络鱼类分类识别模型.Efficient fish-classification is the basis for intelligent monitoring of marine ranching.On one hand,traditional classification methods based on target features through shallow models are inefficient and have poor generalization,which makes it difficult to realize intelligent applications.On the other hand,self-built deep convolutional neural networks(DCNN)model occupies huge computer resources.In view of the above,this paper proposes a method based on DCNN and transfer learning.For the new image-data-set,the pre-training model is retrained by selecting training parameters to realize fish classification and recognition.Experiments show that this method can achieve 97.14%in terms of verification accuracy with a small amount of computer resources.A deep neural network fish classification model with good performance can be obtained by using the learning strategy based on DCNN and parameter transfer.

关 键 词：鱼类分类 深度卷积神经网络 迁移学习 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]