基于水面红外图像的深海网箱鱼群夜间智能监测方法研究  

Research on nighttime intelligent monitoring method for deep-sea cage fish school based on water surface infrared images

作　　者：要紫丹 黄小华[2,3] 李根 胡昱 庞国良[2,3] 袁太平 YAO Zidan;HUANG Xiaohua;LI Gen;HU Yu;PANG Guoliang;YUAN Taiping(Zhejiang Ocean University,Zhoushan 316022,China;South China Sea Fisheries Research Institute,Chinese Academy of Fishery Sciences/Key Laboratory for Sustainable Utilization of Open-Sea Fishery,Ministry of Agriculture and Rural Affairs/Guangdong Cage Engineering Research Center,Guangzhou 510300,China;Research and Development Center for Tropical Aquatic Products,South China Sea Fisheries Research Institute,Chinese Academy of Fishery Sciences/Sanya Tropical Fisheries Research Institute,Sanya 572018,China)

机构地区：[1]浙江海洋大学,浙江舟山316022 [2]中国水产科学研究院南海水产研究所/农业农村部外海渔业可持续利用重点实验室/广东省网箱工程技术研究中心,广东广州510300 [3]中国水产科学研究院南海水产研究所热带水产研究开发中心/三亚热带水产研究院,海南三亚572018

出　　处：《南方水产科学》2024年第1期81-88,共8页South China Fisheries Science

基　　金：国家自然科学基金面上项目(32173024);海南省科技专项资助(ZDYF2021XDYN305,ZDYF2023XDNY066);中国水产科学研究院中央级公益性科研院所基本科研业务费专项资金资助(2023TD97);中国水产科学研究院南海水产研究所中央级公益科性研院所基本科研业务费专项资金资助(2023RC01,2022TS06);农业农村部外海渔业可持续利用重点实验室开放基金(LOF 2023-01)。

摘　　要：获取深海网箱养殖过程鱼群活动数据,开展鱼群监测是提升深海养殖效率、降低养殖成本的有效手段。基于水面红外相机,利用深度学习前沿技术,提出了一种鱼群智能监测方法。该方法涉及鱼群识别及计数、鱼体分割和鱼体游向判断3个功能模块。首先,通过红外相机采集鱼类的图像信息,并进行标注以构建数据集,然后采用改进的Faster RCNN模型,以Mobilenetv2+FPN网络作为特征提取器,实现鱼类的准确识别,并输出包围框表征鱼类个体位置。其次,从框图内选择亮度前20%的像素点作为分割提示点,利用Segment Anything Model对图像进行分割,生成鱼体分割图。最后,通过对鱼体分割图进行椭圆拟合处理,可以判定鱼类的游向信息。改进的Faster RCNN模型在进行100次迭代训练后,平均精确率达到84.5%,每张图片的检测时间为0.042 s。结果表明,在水面红外图像的鱼类数据集上,所提出的改进Faster RCNN模型和椭圆拟合等关键技术能够实现对鱼群的自动监测。Obtaining fish school information on its size and behavior through fish school monitoring is an important way to improve the efficiency of deep sea aquaculture and reduce costs.In this study,an intelligent fish school monitoring method is proposed by using infrared cameras mounted on a net cage for data collection,in addition to the latest deep learning techniques for model training.The method involves three functional modules:fish detection,fish segmentation and fish pose determination.Firstly,fish images were collected by infrared cameras and manually annotated to build datasets,while an improved faster RCNN model that uses Mobilenetv2 and FPN network as feature extractors to improve detection accuracy is adopted to output bounding boxes of individual fish.Secondly,the top 20%of brightness pixels in the block map were selected as segmentation prompt points,and the image was segmented using Segment Anything Model to generate fish segmentation results.Finally,the fish pose information was determined by applying elliptical fitting using fish segmentation results.After 100 epochs of training,the average precision(AP)of the improved Faster RCNN model reached 84.5%,and the detection time per image was 0.042 s.The results indicate that the proposed method can achieve automatic monitoring of fish school on infrared images and extract effective information.

关 键 词：深海网箱 鱼群监测 红外图像 目标检测 实例分割 

分 类 号：S951.2[农业科学—水产养殖]