[Paper Review] FinSight-Net:A Physics-Aware Decoupled Network with Frequency-Domain Compensation for Underwater Fish Detection in Smart Aquaculture

Jinsong Yang, Zeyuan Hu+2 more|arXiv (Cornell University)|Feb 23, 2026

Water Quality Monitoring Technologies0 citations

TL;DR

FinSight-Net introduces physics-aware MS-DDSP bottlenecks and an EPA-FPN neck to mitigate wavelength-dependent absorption and backscattering in underwater fish detection, achieving state-of-the-art accuracy with fewer parameters.

ABSTRACT

Underwater fish detection (UFD) is a core capability for smart aquaculture and marine ecological monitoring. While recent detectors improve accuracy by stacking feature extractors or introducing heavy attention modules, they often incur substantial computational overhead and, more importantly, neglect the physics that fundamentally limits UFD: wavelength-dependent absorption and turbidity-induced scattering significantly degrade contrast, blur fine structures, and introduce backscattering noise, leading to unreliable localization and recognition. To address these challenges, we propose FinSight-Net, an efficient and physics-aware detection framework tailored for complex aquaculture environments. FinSight-Net introduces a Multi-Scale Decoupled Dual-Stream Processing (MS-DDSP) bottleneck that explicitly targets frequency-specific information loss via heterogeneous convolutional branches, suppressing backscattering artifacts while compensating distorted biological cues through scale-aware and channel-weighted pathways. We further design an Efficient Path Aggregation FPN (EPA-FPN) as a detail-filling mechanism: it restores high-frequency spatial information typically attenuated in deep layers by establishing long-range skip connections and pruning redundant fusion routes, enabling robust detection of non-rigid fish targets under severe blur and turbidity. Extensive experiments on DeepFish, AquaFishSet, and our challenging UW-BlurredFish benchmark demonstrate that FinSight-Net achieves state-of-the-art performance. In particular, on UW-BlurredFish, FinSight-Net reaches 92.8% mAP, outperforming YOLOv11s by 4.8% while reducing parameters by 29.0%, providing a strong and lightweight solution for real-time automated monitoring in smart aquaculture.

Motivation & Objective

• Address the degradation of underwater imagery due to absorption and scattering in smart aquaculture.
• Develop a lightweight, physics-aware detector that decouples feature streams to restore biological cues.
• Improve multi-scale feature fusion to preserve high-frequency details in turbid water.

Proposed method

• Proposes Multi-Scale Decoupled Dual-Stream Processing (MS-DDSP) bottleneck with four parallel branches for frequency-domain compensation.
• Implements Efficient Path Aggregation FPN (EPA-FPN) with vertical long-range skip connections and path pruning.
• Uses channel-wise soft attention to weight branch outputs based on global statistics.
• Backbone is CSPDarknet integrated with EPA-FPN and MS-DDSP for robust feature extraction.
• Grounds the design in the Jaffe–McGlamery optical model to separate backscatter and absorption effects.

Experimental results

Research questions

• RQ1 Can a physics-aware, decoupled network outperform black-box CNN/Transformer detectors for underwater fish detection?
• RQ2 Does a lightweight MS-DDSP+EPA-FPN architecture maintain high accuracy under turbidity and occlusions while reducing parameters?
• RQ3 How do frequency-domain compensation and detail-preserving fusion affect localization of non-rigid fish in turbid water?

Key findings

• FinSight-Net achieves state-of-the-art mAP50 across three underwater datasets (DeepFish, AquaFishSet, UW-BlurredFish).
• On UW-BlurredFish, FinSight-Net reaches 92.8% mAP, outperforming YOLOv11s by 4.8% with 29% fewer parameters.
• Ablations show EPA-FPN boosts mAP50 by 6.0% with reduced parameters, and MS-DDSP provides substantial gains via physics-aware branches.
• Full model with EPA-FPN and MS-DDSP attains 53.4% mAP50 on UW-BlurredFish with 6.7M parameters and 20.4 GFLOPs.
• Generalization tests: trained on DeepFish, evaluated on UW-BlurredFish without fine-tuning, demonstrating robustness to turbidity and lighting variations.