基于L(2+1)D的养殖鱼类摄食状态下活跃程度识别方法  

Method for identifying activity level of cultivated fish under feeding state based on L(2+1)D

作　　者：唐晓萌 缪新颖 TANG Xiaomeng;MIAO Xinying(Information Engineering College,Dalian Ocean University,Dalian 116023,China;Key Laboratory of Environment Controlled Aquaculture,Ministry of Education,Dalian 116023,China)

机构地区：[1]大连海洋大学信息工程学院,辽宁大连116023 [2]设施渔业教育部重点实验室,辽宁大连116023

出　　处：《现代电子技术》2024年第8期155-159,共5页Modern Electronics Technique

基　　金：设施渔业教育部重点实验室基金。

摘　　要：鱼类行为的活跃程度是鱼类行为研究中的关键指标,可为水产养殖过程提供有用的基础数据。然而现有的计算机视觉方法在活跃程度识别的应用中依赖于大量存储和计算资源,在实际场景中实用性较差。为了解决这些问题,提出一种鱼类摄食活动识别模型——L(2+1)D,将3D卷积分解为2D大空间卷积和1D时间卷积,使用少量的大型卷积核来增加感受野,实现更强大的特征提取效果。将空间卷积和时间卷积串联成用于时空特征学习的时空模块,并减少时空模块数量,达到减少参数数量的同时提高准确性的效果。实验结果表明,所提方法可以在实际水产养殖中准确识别鱼群的活跃程度,准确率可达到65.02%,并更适合部署在资源受限的设备或现场。The activity level of fish behavior is a key indicator in the fish behavior research,providing useful basic data for aquaculture processes.The existing methods based on computer vision rely on a large amount of storage and computing resources in the application of activity level recognition,which has poor practicality in practical scenarios.To address these limitations,a fish feeding activity recognition model named L(2+1)D is proposed.3D convolution is decomposed into 2D large spatial convolution and 1D temporal convolution.A small number of large convolutional kernels are used to increase receptive fields,so as to realize a more powerful feature extraction effect.The spatial convolution and temporal convolution are concatenated into spatiotemporal modules for feature extraction,and the number of spatiotemporal modules is reduced,achieve the effect of reducing the number of parameters while improving accuracy.The experimental results show that the proposed method can accurately identify the activity level of fish schools in actual aquaculture,with an accuracy rate of 65.02%,and is more suitable for deployment in resource limited equipment or sites.

关 键 词：鱼类活跃程度 卷积神经网络 图像预处理 特征提取 时空特征学习 行为量化 

分 类 号：TN957.52-34[电子电信—信号与信息处理]