融合摄食过程声像特征的鱼类摄食强度量化方法研究  

SENSITIVE DETECTION OF FISH FEEDING INTENSITY BY USING SOUND AND IMAGE FEATURES OF FEEDING PROCESS

作　　者：郑金存 叶章颖[2] 赵建[2] 张慧 黄平 覃斌毅 庞毅 ZHENG Jin-Cun;YE Zhang-Ying;ZHAO Jian;ZHANG Hui;HUANG Ping;QIN Bin-Yi;PANG Yi(College of Physics and Telecommunication Engineering,Yulin Normal University,Yulin 537000,China;College of Biosystems Engineering and Food Science,Zhejiang University,Hangzhou 310058,China;College of Biology and Pharmacy,Yulin Normal University,Yulin 537000,China;Guangxi Colleges and Universities Key Lab of Complex System Optimization and Big Data Processing,Yulin 537000,China)

机构地区：[1]玉林师范学院物理与电信工程学院,广西玉林537000 [2]浙江大学生物系统工程与食品科学学院,浙江杭州310058 [3]玉林师范学院生物与制药学院,广西玉林537000 [4]广西高校复杂系统优化与大数据处理重点实验室,广西玉林537000

出　　处：《海洋与湖沼》2024年第3期577-588,共12页Oceanologia Et Limnologia Sinica

基　　金：国家大宗淡水鱼产业技术体系项目,CARS-45-24号;广西重点研发计划项目,2022AB20139号;国家级大学生创新创业训练计划项目,202210606017号。

摘　　要：基于鱼类摄食行为反馈的精准投喂是确保饲料高效利用与降低水体污染的有效手段,针对当前单一传感器难以精确测量鱼群摄食强度的难题,提出一种基于摄食过程声像特征融合的鱼类摄食强度量化方法。首先利用深度图包含的三维空间信息分析水体表层摄食鱼类数量,设计基于帧间差分运算的深度图能量变化测量系统表征鱼群摄食活跃程度;进而利用近红外光源因水面反射而导致的高亮度饱和点在近红外图中的位置变化测量水体流场的波动程度;同时利用音轨记录仪存储摄食音频。最后通过加权融合方式,综合摄食动态、水体流场变化、摄食音频三类具有不同物理属性的特征信息,精确量化了鱼群摄食强度,总体识别精确度达到97%。本文采用新的成像技术,取得分析速度与分析精度的最佳平衡,为精准投喂提供了一种鲁棒性强、分析速度快的实用方法。Accurate assessment of fish appetite is of great significance for guiding feeding and production practice.However,most of the previous methods for assessment of fish feeding intensity have problems of high computational complexity and low precision.To achieve an automatic objective evaluation of fish feeding intensity,we proposed an improved multi-features fusion algorithm.First,a multi-track audio recorder was used to store the feeding audio.Due to the positive correlation between feeding audio and feeding intensity,the amplitude of the audio was as one of the parameters to evaluate feeding intensity.Secondly,a depth camera named Azure Kinect was selected to analyze the feeding dynamics of fish.The sensor based on 3D single point imaging technology,which can provide geometric information of 3D environment with high-frame rate.The whole feeding process was recorded in real time by using depth image and near infrared image,and 3D spatial information contained in the depth map was used to count the number of fish at the surface of water.During fish preying at water surface,the pixel value of the depth map fluctuates strongly.The feeding intensity of fish was quantified by continuously calculating the difference of depth map.Thirdly,the intensity of water level fluctuation during feeding was measured synchronously by using an IR camera,which is included in Azure Kinect system.Fluctuation of water caused by feeding changes the condition of specular reflection,the position of high brightness pixels produced by the reflection of near-infrared light source changes significantly in a near-infrared image.Finally,by combining the characteristic information of feeding dynamics,water level fluctuation,and feeding audio,the evaluation accuracy of fish feeding intensity reached 97%.This study adopted new imaging techniques to achieve the best performance in analysis speed and accuracy,and provided a practical method in strong robustness and fast analysis speed for precise feeding.

关 键 词：鱼摄食强度 近红外图 深度图 摄食音频 加权融合 

分 类 号：S951.2[农业科学—水产养殖] TP391.4[农业科学—水产科学]