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

Jinsong Yang; Zeyuan Hu; Yichen Li; Hong Yu

cs.CVFeb 23, 2026

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

Authors: Jinsong Yang, Zeyuan Hu, Yichen Li, Hong Yu

Organizations: Dalian Key Laboratory of Smart Fishery, Dalian, China · College of Information Engineering, Dalian Ocean University, China · Liaoning Provincial Key of Marine Information Technology, China · Key Laboratory of Environment Controlled Aquaculture, Dalian, China ·

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.

Figures & tables

Figure 1: Each method’s overall performance(’P’) on test data and generalization performance (’G’) on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics.

Figure 2: Architectural topology and physics-aware paradigm of FinSight-Net. (Top) Overall pipeline integrating CSPDarknet with EPA-FPN (Module A) and MS-DDSP (Module B) for robust feature extraction. (Bottom) Conceptual framework mapping Jaffe-McGlamery optical constraints to recovery streams: High-Frequency Recovery suppresses backscattering ( $I_{bs}$ ), while Spectral Restoration compensates for wavelength-dependent absorption ( $e^{-\eta d}$ ).

Figure 3: MS-DDSP Bottleneck. This module employs parallel heterogeneous branches to decouple and compensate for underwater optical degradations , utilizing soft attention to optimize SNR and restore biological structural fidelity.

Figure 4: EPA-FPN Architecture. This detail-filling mechanism utilizes vertical long-range skip connections to inject high-resolution spatial cues into deep semantic nodes, salvaging critical high-frequency structural information.

Model	Param.	FLOPs	DeepFish					AquaFishSet				UW-BlurredFish
Model	(M)	(G)	F1	P	R	mAP 50	mAP 50-95	P	R	mAP 50	mAP 50-95	P	R	mAP 50	mAP 50-95
Faster R-CNN	41.3	138.5	79.5	83.4	75.8	83.4	50.4	85.3	80.2	86.4	53.6	82.4	73.6	82.4	50.4
Mask R-CNN	43.9	188.4	81.2	85.6	77.2	84.2	51.2	85.8	80.6	86.9	53.2	83.4	75.2	82.2	50.2
CAM-RCNN	44.0	189.0	81.5	86.2	77.4	84.8	51.4	86.0	80.8	87.0	53.4	83.2	77.2	83.0	50.5
YOLO-Fish	61.8	65.9	84.1	86.8	82.2	88.6	52.2	85.5	80.4	84.9	52.2	82.7	78.4	82.9	50.6
CEH-YOLO	4.4	11.6	83.6	87.2	80.2	87.5	50.2	89.0	82.8	87.2	55.2	87.2	85.0	88.9	53.2

Table 1: Quantitative comparison with state-of-the-art object detectors. Bold and underline indicate the best and second-best performance, respectively. All models are evaluated at a

640\times 640

resolution. Our FinSight-Net achieves state-of-the-art results across all three benchmark datasets while maintaining competitive efficiency.

Configuration	EPA-FPN	MS-DDSP	mAP 50	Params	Lat. (ms)
Baseline (CSPDarknet+PANet)			44.3	9.7M	12.4
+ EPA-FPN	✓		50.3	6.5M	8.6
+ MS-DDSP		✓	46.8	9.9M	12.6
FinSight-Net (Full)	✓	✓	53.4	6.7M	8.6

Table 2: Ablation study of FinSight-Net components on the UW-BlurredFish dataset.

\checkmark

denotes the inclusion of the corresponding module. Bold indicates the best performance.

Figure 5: Qualitative comparison of detection results. Compared to the YOLOv11s baseline, FinSight-Net significantly reduces missed detections and yields higher confidence scores in turbid and densely occluded aquaculture scenarios.

Neck Architecture	mAP 50	Params (M)	FLOPs (G)	Lat. (ms)
Top-down FPN	44.8	1.2	11.4	12.4
Bi-FPN	47.5	1.4	13.5	8.6
PANet	50.2	2.3	28.8	13.2
EPA-FPN (Ours)	50.3	1.6 †	15.2	8.6
† Denotes a significant 30% parameter reduction compared to the PANet baseline.

Table 3: Ablation study of different Neck architectures on the UW-BlurredFish dataset. Bold and underline indicate the best and second-best results, respectively. Our EPA-FPN achieves a superior balance between accuracy and computational cost.

Configuration	mAP 50 (%)	$\Delta$
Full MS-DDSP (B1+B2+B3+B4)	90.4	—
w/o Branch 2 (Turbidity Suppression)	87.2	$-3.2$
w/o Branch 3 (Absorption Compensation)	88.5	$-1.9$
w/o Branch 4 (Detail Preservation)	89.1	$-1.3$

Table 4: Ablation analysis of MS-DDSP branch contributions on the UW-BlurredFish dataset. Bold indicates the performance of our full configuration. The removal of any physics-aware branch leads to consistent performance degradation.

Figure 6: Feature Power Spectrum Analysis. Underwater backscattering causes significant high-frequency energy decay in the baseline (YOLOv11s). FinSight-Net effectively salvages these structural cues through EPA-FPN’s long-range skip connections , ensuring superior fidelity and precise localization in turbid environments.

Figure 7: Adaptive Branch Activation Weights.Interpretability of the MS-DDSP bottleneck. The radar plot demonstrates dynamic branch re-weighting via soft attention : B2 (decorrelation) is prioritized to suppress turbidity-induced backscattering , while B3 (spectral weighting) compensates for wavelength-dependent absorption during spectral distortion.

Figure 8: Feature map visualization comparison. FinSight-Net enhances SNR via physics-aware MS-DDSP and EPA-FPN , effectively suppressing backscattering while salvaging high-frequency structural details to ensure precise fish-background discrimination in turbid environments.

Underwater Object Detection

Underwater object detection localizes and classifies targets in degraded camera and sonar data for marine science, robotics, and inspection.

See all 28 papers

Adaptive Enhancement and Dual-Pooling Sequential Attention for Lightweight Underwater Object Detection with YOLOv10
Mar 4, 2026Md. Mushibur Rahman, Umme Fawzia Rahim, Enam Ahmed TaufikObject Detection Underwater Object Detection
Department of Computer Science and Engineering, Dhaka University of Engineering & Technology Gazipur, 1707, Bangladesh · Department of Computer Science and Engineering, BRAC University Dhaka, Bangladesh
AQUA-Net: Adaptive Frequency Fusion and Illumination Aware Network for Underwater Image Enhancement
Dec 5, 2025Munsif Ali, Najmul Hassan, Lucia Ventura +3Image Enhancement Underwater Image Enhancement
Stazione Zoologica Anton Dohrn, Villa Comunale, 80121 Napoli, Italy. · School of Computer Science and Engineering, The University of Aizu, Aizuwakamatsu, Japan.
SPMamba-YOLO: An Underwater Object Detection Network Based on Multi-Scale Feature Enhancement and Global Context Modeling
Feb 26, 2026Guanghao Liao, Zhen Liu, Liyuan Cao +2Underwater Object Detection Efficient Object Detection
School of Electronic and Information Engineering, University of Science and Technology Liaoning, Anshan, 114051, China

Mar 4, 2026J/K move · Enter open · S save

Adaptive Enhancement and Dual-Pooling Sequential Attention for Lightweight Underwater Object Detection with YOLOv10

Md. Mushibur Rahman, Umme Fawzia Rahim, Enam Ahmed Taufik

Department of Computer Science and Engineering, Dhaka University of Engineering & Technology Gazipur, 1707, Bangladesh · Department of Computer Science and Engineering, BRAC University Dhaka, Bangladesh

Object Detection Underwater Object Detection Efficient Object Detection Attention Mechanisms

Open paper AskPDF arXiv

Underwater object detection constitutes a pivotal endeavor within the realms of marine surveillance and autonomous underwater systems; however, it presents significant challenges due to pronounced visual impairments arising from phenomena such as light absorption, scattering, and diminished contrast. In response to these formidable challenges, this manuscript introduces a streamlined yet robust framework for underwater object detection, grounded in the YOLOv10 architecture. The proposed method integrates a Multi-Stage Adaptive Enhancement module to improve image quality, a Dual-Pooling Sequential Attention (DPSA) mechanism embedded into the backbone to strengthen multi-scale feature representation, and a Focal Generalized IoU Objectness (FGIoU) loss to jointly improve localization accuracy and objectness prediction under class imbalance. Comprehensive experimental evaluations conducted on the RUOD and DUO benchmark datasets substantiate that the proposed DPSA_FGIoU_YOLOv10n attains exceptional performance, achieving mean Average Precision (mAP) scores of 88.9% and 88.0% at IoU threshold 0.5, respectively. In comparison to the baseline YOLOv10n, this represents enhancements of 6.7% for RUOD and 6.2% for DUO, all while preserving a compact model architecture comprising merely 2.8M parameters. These findings validate that the proposed framework establishes an efficacious equilibrium among accuracy, robustness, and real-time operational efficiency, making it suitable for deployment in resource-constrained underwater settings.

Abstract

Underwater Object Detection

Explore similar work

Adaptive Enhancement and Dual-Pooling Sequential Attention for Lightweight Underwater Object Detection with YOLOv10

AQUA-Net: Adaptive Frequency Fusion and Illumination Aware Network for Underwater Image Enhancement

SPMamba-YOLO: An Underwater Object Detection Network Based on Multi-Scale Feature Enhancement and Global Context Modeling