FishMP: multi-scale parallel perception for efficient underwater fish detection

ABSTRACT Underwater fish detection faces significant challenges due to lighting variations, water turbidity, and scale differences. This paper proposes a novel fish detection model, FishMP, which integrates an efficient Multi-Scale Parallel Perception Module (MPPM) with a P2 multi-branch cascaded sequence fusion network. The MPPM employs a multi-branch feature extraction strategy and attention mechanisms to capture features at different levels, effectively preventing the loss of critical information during downsampling and improving detection accuracy for blurred targets. To mitigate information loss during feature integration, the P2 multi-branch network is optimized using triple-feature serial modules and an ordered fusion pathway. Furthermore, an additional P2 low-level spatial information branch is introduced to enrich fish morphological and spatial position features, ensuring robust performance in complex underwater environments. Experimental results on the underwater datasets DeepFish and Fishway demonstrate that FishMP achieves an outstanding balance between speed and accuracy.