改进残差网络的海水养殖鱼类识别与分类研究  

Improved residual network and its application in intelligent mariculture

作　　者：季星宇 赵雪峰 陈荣军 仲兆满 Ji Xingyu;Zhao Xuefeng;Chen Rongjun;Zhong Zhaoman(College of Computer Engineering,Jiangsu Ocean University,Lianyungang,Jiangsu 222000,China)

机构地区：[1]江苏海洋大学计算机工程学院,江苏连云港222000

出　　处：《计算机时代》2023年第9期101-105,共5页Computer Era

基　　金：国家自然科学基金(72174079);江苏省苏北科技专项(SZ-LYG202024);江苏省“青蓝工程”优秀教学团队(2022-29)。

摘　　要：为了满足海水养殖行业不断提高的智能化需求,对海洋鱼类的识别和分类算法进行研究。采用多重残差网络进行鱼类识别及分类,不仅降低计算复杂度,同时加快了残差网络的学习速度;引入指数线性单元(ELU)改进网络的标准残差模块,对输入的负激活值部分进行非线性变化,其参数可通过卷积训练进行自适应学习,同时保持正激活值部分不变,解决了传统残差模块中ReLU层将包含有用信息的负激活值完全丢弃的问题,以降低梯度消失的概率。在海洋鱼类识别与分类的多次实验中,改进的残差网络准确率均不低于95.48%,表明改进算法拥有较高的识别准确率和良好的稳定性。In order to meet the increasing intelligent needs of mariculture industry,the recognition and classification algorithms of marine fish are studied.A multiple residual network is used for fish recognition and classification.This network not only reduces computational complexity,but also accelerates the learning speed of the residual network.Exponential linear unit(ELU)is introduced to improve the standard residual module of the network,which performs nonlinear changes on the negative activation value portion of the input.Its parameters can be adaptively learned through convolution training,while maintaining the positive activation value portion unchanged.It solves the problem of completely discarding the negative activation value containing useful information in the ReLU layer of the traditional residual module,and reduces the probability of gradient disappearance.In many experiments on marine fish recognition and classification,the accuracy of the improved residual network has reached 95.48%or more,which indicates that the improved algorithm has a high recognition accuracy and good stability.

关 键 词：海水养殖 鱼类识别 残差网络 指数线性单元 激活函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]