文献类型：期刊文献

英文题名：Segmentation and calculation of splashes area during fish feeding using deep learning and linear regression

作者：Qian, Liwen[1,2];Yu, Guoyan[1,2];Liu, Haochun[1,2];He, Zijian[1,2]

机构：[1]Guangdong Ocean Univ, Sch Mech Engn, Zhanjiang 524088, Guangdong, Peoples R China;[2]Guangdong Ocean Univ, Guangdong Marine Equipment & Mfg Engn Technol Res, Zhanjiang 524088, Guangdong, Peoples R China

年份：2025

卷号：234

外文期刊名：COMPUTERS AND ELECTRONICS IN AGRICULTURE

收录：SCI-EXPANDED(收录号：WOS:001467453400001)、、EI(收录号：20251218099500)、Scopus(收录号：2-s2.0-105000468213)、WOS

基金：This research was supported by the Scientific Research Start-Up Funds of Guangdong Ocean University (060302062201) , Guangdong Agricultural Technology Service Light Cavalry Major Agricultural Technology Rural Promotion Project (NJTG20240240) , Fund of Guangdong Provincial Key Laboratory of Intelligent Equipment for South China Sea Marine Ranching (2023B1212030003) , the Innovation Team Project for Ordinary Universities in Guangdong Province (2024KCXTD041) , Zhanjiang Key Laboratory of Modern Marine Fishery Equipment (2021A05023) and the Guangdong Provincial Graduate Education Innovation Programme Project (2023JGXM_75) .

语种：英文

外文关键词：Feeding intensity of fish; Binocular camera; Deep learning; Linear regression

外文摘要：Evaluating the feeding intensity of fish in large deep-sea cages is crucial for improving feed utilization efficiency and reducing aquaculture costs. During feeding operations, fish engage in intense feeding competition and jumping behaviors, creating splashes of varying sizes that indirectly reflect their hunger intensity. This study focuses on the splashes generated by golden pompano during feeding and introduces a splash area estimation model, SSS-YOLOv8n-seg-DI, which combines linear regression and deep learning, providing foundational data for non-invasive appetite analysis. Firstly, the model employs the StarNet network and a self-designed SCD-Head shared convolutional detection head to make the YOLOv8n-seg network more lightweight, resulting in the SSSYOLOv8n-seg model. This modification improves accuracy while reducing parameters and FLOPs by 53.99 % and 32.50 %, respectively. Secondly, to reduce equipment costs, this paper utilizes a binocular camera and proposes a target area calculation model, DI, based on linear regression, which calculates the target area using the depth information provided by the camera. Ultimately, the combined SSS-YOLOv8n-seg-DI model is capable of segmenting and calculating the splash area generated by fish during feeding, providing foundational data for appetite analysis and adjusting subsequent feeding quantities. Results from marine experiments show that the calculated splash area achieved an R2 of 0.931, with a mean absolute error (MAE) of 0.813 m2 and a root mean square error (RMSE) of 0.869 m2. The experimental outcomes demonstrate that the model exhibits strong robustness, meeting the requirements for splash area calculation in complex environments.