基于MobileViT-CBAM-BiLSTM的开放式养殖环境鱼群摄食强度分类模型  

Classification Model of Fish Feeding Intensity Based on MobileViT-CBAM-BiLSTM

作　　者：徐立鸿 黄志尊[1] 龙伟 蒋林华 童欣[1] XU Lihong;HUANG Zhizun;LONG Wei;JIANG Linhua;TONG Xin(School of Information Engineering,Huzhou University,Huzhou 313000,China;College of Electronics and Information Engineering,Tongji University,Shanghai 201804,China)

机构地区：[1]湖州师范学院信息工程学院,湖州313000 [2]同济大学电子与信息工程学院,上海201804

出　　处：《农业机械学报》2024年第11期147-153,共7页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(62373286、62175037);湖州市重点研发计划农业“双强”专项(2022ZD2060)。

摘　　要：鱼群摄食的精准投喂技术是实现智慧化水产养殖的一项关键技术。大多数精准投喂模型都是基于水质较清晰的室内养殖池,不太适用于开放式养殖环境。本研究通过水上视角采集构建了一套开放式池塘数据集,并对数据集进行数据增强增加其多样性,然后在轻量化神经网络MobileViT基础上,将CBAM注意力模块与MV2模块结合设计了CBAM-MV2模块,并嵌入BiLSTM循环神经网络用于识别分类,提出改进的MobileViT-CBAM-BiLSTM模型,提高了模型预测能力、鲁棒性和泛化性能,实现了鱼群摄食行为的三分类。实验结果显示,改进后MobileViT在采集的视频帧数据集上明显优于改进前的MobileViT,准确率98.61%,宏F1值达98.79%,相对于原始MobileViT准确率提高6.33个百分点,宏F1值提高6.75个百分点。Precise feeding technology for fish ingestion is a key technology to achieve intelligent aquaculture. However, most of the precise feeding model is based on indoor aquaculture ponds with clear water quality, which are not suitable for outdoor open farming environments. In view of the actual situation, a set of detailed open pond dataset through water perspective acquisition was constructed, and the dataset was augmented to increase its diversity, and then the BiLSTM bidirectional recurrent neural network was embeded on the basis of the lightweight neural network MobileViT, so as to improve the memory ability of the model for video sequence data in a long period of time, and the CBAM attention module was combined with the MV2 module to design the CBAM-MV2 module, and then the CBAM-MV2 module was added to different layers of the model for experiments to obtain the most reasonable improvement scheme. Finally, an improved MobileViT-CBAM-BiLSTM fish feeding behavior classification model was proposed, which improved the prediction ability, robustness and generalization performance of the model, and realized the three classification of fish feeding behavior. The experimental results showed that the improved MobileViT was significantly better than previous in the collected video frame dataset, with an accuracy of 98.61%, 98.79% for Macro-F1, which was 6.33 percentage points for accuracy, 6.75 percentage points for Macro-F1 compared with the original MobileViT.

关 键 词：鱼群摄食强度分类模型 精准投喂 MobileViT BiLSTM CBAM 

分 类 号：S951.2[农业科学—水产养殖]