基于深度学习的鱼类智能识别系统的设计与实现  

Design of fish intelligent recognition system based on deep learning

作　　者：吕俊霖[1] 麦嘉铭 熊浩 蔡海真 LYU Junlin;MAI Jiamin;Xiong Hao;Cai Haizhen(South China Fisheries Research Institute,CAFS,GuangZhou 510300,China;South China Agricultural University,GuangZhou 510642,China)

机构地区：[1]中国水产科学研究院南海水产研究所,广州510300 [2]华南农业大学,广州510642

出　　处：《渔业现代化》2021年第3期90-96,共7页Fishery Modernization

基　　金：农业农村部财政专项项目(NFZX2021);中国水产科学研究院南海水产研究所中央级公益性科研院所基本科研业务费专项资金(2019TS23)。

摘　　要：中国拥有种类繁多的鱼类,外形是其分类的重要依据。但目前主要采用人工识别方法进行分类,为解决鱼类人工识别存在的问题,提出一种基于深度学习的鱼类智能识别系统的设计,以实现对中国1400种鱼类的智能识别。系统首先采用卷积神经网络的Efficient模型,将含有1400种鱼类,50万张鱼类图片的数据集进行训练,最终得到的模型识别精度达到了95%,单张图片识别时间仅为0.2 s,模型大小为74.5 MB。系统前端使用微信小程序,后端采用Spring+SpringMVC+Mybatis的SSM架构,调用训练好的模型文件进行识别,实现了鱼类识别、页面呈现、统计分析和相邻种类推荐等功能。该系统所提出的设计和实现方法对鱼类智能识别技术在移动端的应用提供了一种可行的思路,对海洋科研人员和开发人员有一定的借鉴作用。There are many kinds of fish in China,and their shape is an important basis for their classification.In order to solve the difficulty in artificial fish identification,an intelligent fish identification system based on deep learning has been proposed to realize the intelligent identification of 1400 species of fish in China.The system first uses convolutional neural network’s EfficientNet model to train a dataset of 500,000 pictures of 1,400 species of fish.The accuracy of the model is 95%,and the recognition time of a single picture is only 0.2 seconds,the model’s size is 74.5 MB.Then the front-end of the system uses Wechat applet,the back-end uses Spring+Spring MVC+Mybatis SSM architecture,invoking the training of the model file for identification,the realization of fish identification,page rendering,statistical analysis and recommendation of adjacent species.The system’s design and implementation provide a feasible idea for the application of intelligent fish identification technology to mobile terminal,which marine researchers and developers can use as reference.

关 键 词：鱼类 智能识别 深度学习 EfficientNet 设计与实现 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]