Fish density estimation for smart marine aquaculture using a multi-granularity counting consistency network

Abstract Fish density estimation is an important task in precision aquaculture, supporting feeding management, stocking regulation, and production monitoring. However, accurate estimation in aquaculture scenes remains challenging due to severe occlusion, lighting variation, water turbidity, and background interference. Existing fish density estimation methods mainly focus on image-level count accuracy or pixel-level density-map supervision but often lack explicit constraints on local count consistency, which may lead to inaccurate spatial distributions in dense scenes. Furthermore, features learned mainly from visual appearance are easily affected by complex visual backgrounds in aquaculture scenes. To address these issues, we propose MGCC-Net, a multi-granularity counting consistency network for fish density estimation in aquaculture scenes. MGCC-Net introduces two components: a multi-granularity local patch count loss to enforce local count consistency across multiple spatial scales and a dual-projection channel attention module that generates channel weights from two learned projections of the global visual descriptor and is jointly optimized with the density-estimation objective. We also construct AquaFish-Count, a real-world fish density dataset containing images from aquaculture and natural marine environments, organizing it into FD-High and FD-Low based on the annotated fish count in each image. Experimental results show that MGCC-Net achieves the best performance on FD-High and remains competitive on FD-Low, with its clearest advantage observed in dense scenes with severe fish overlap. These results demonstrate the effectiveness of MGCC-Net for fish density estimation and its potential application in marine biological environment perception and smart aquaculture monitoring systems.