Work

SLDCNet: Shallow Latent Density Collective Representation for Underwater Fish School Abnormality Recognition

Year: 2026
Type: conference-paper
Abstract: Aquaculture cage environments commonly exhibit high-density fish distributions and severe occlusion, limiting the reliability of instance-level detection and tracking for abnormal behavior monitoring. To address this challenge, a density-driven framework for underwater fish school abnormality recognition is introduced, in which fish schools are represented by shallow collective latent densities rather than individual trajectories. The framework follows a two-stage learning paradigm. A Shallow Latent Density Collective Network (SLDCNet) is employed to generate fine-grained density maps that encode spatial distribution patterns as compact collective-state representations under complex underwater conditions. The resulting density representations are subsequently processed by a factorized $\mathbf{3} \times \mathbf{1 D}$ convolutional architecture to capture spatiotemporal behavioral transitions and identify abnormal dynamics. This representation-centric formulation enables characterization of abnormal behavior without explicit fish detection or tracking. Experimental results indicate that the proposed method achieved good performance with 97.85% accuracy, 99.96% precision, 95.70% recall, and 97.80% F1-score for fish school behavior recognition, and it demonstrated applicability as an effective tool for cage aquaculture monitoring.
Language: English
Open Access status: hybrid
OpenAlex ID: W7204804685
FWCI:
Cites: 6
Cited by:
Related to: 10