文献类型：期刊文献

英文题名：Aquaculture fish counting via joint training of density estimation and multi-scale detection

作者：Ji, Ziliang[1]; Lin, Cong[2,4]; Rachel Merveille, Fomekong Fomekong[1]; Hou, Mingxin[3]; Liu, Mingxin[2,4]

机构：[1] College of Naval Architecture and Shipping, Guangdong Ocean University, Zhanjiang, 524088, China; [2] College of Electronic and Information Engineering, Guangdong Ocean University, Zhanjiang, 524088, China; [3] School of Mechanical Engineering, Guangdong Ocean University, Zhanjiang, 524088, China; [4] Guangdong Provincial Key Laboratory of Intelligent Equipment for South China Sea Marine Ranching, Zhanjiang, 524088, China

年份：2026

卷号：199

外文期刊名：Applied Soft Computing

收录：EI(收录号：20261920685150)

语种：英文

外文关键词：Errors - Fish - Fisheries - Learning systems - Marine applications - Signal encoding

外文摘要：Accurate fish counting is essential for precision aquaculture. However, in fence-based marine ranching, complex open-water backgrounds, body occlusion, and scale variation make this task particularly challenging. In this paper, we propose a joint-training-based multi-scale perception and counting framework (JMPCNet) tailored to aquaculture applications. JMPCNet couples a scale-aware density regressor with an auxiliary detector in a mutually reinforcing manner. The density branch employs a multi-scale receptive field fusion module (MRFFM) to fuse multi-stage backbone features and produce density maps that encode global spatial distribution and scale cues. These density maps are then injected as spatial priors into the SAM-DETR++ encoder to modulate multi-scale features learning, improving discrimination under cluttered backgrounds and heavy occlusion. Meanwhile, instance-level detection supervision regularizes the shared representation, alleviating the ambiguity of density regression when foreground and background are visually similar. Through joint optimization of distribution-level and instance-level objectives, JMPCNet achieves improved counting performance in dense aquaculture environments. Experiments on the IOCfish5K and YoutubeFish-35 datasets demonstrate consistent performance gains across multiple evaluation metrics. On YoutubeFish-35, JMPCNet reduces the mean absolute error (MAE), mean squared error (MSE), and normalized absolute error (NAE) from 44.61, 79.81, and 2.52 to 40.25, 71.56, and 2.36, respectively. On the more challenging IOCfish5K test set, it further decreases these metrics from 17.12, 41.25, and 0.38 to 16.97, 38.71, and 0.36, respectively. Additional experiments on the Fish4Knowledge dataset verify its effectiveness under degraded underwater imaging conditions, while further generalization studies on two high-density crowd-counting datasets suggest its broader applicability to dense counting tasks. ? 2026 Elsevier B.V.