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

FinSight-Net is an efficient, physics-aware underwater fish detection framework that leverages a Multi-Scale Decoupled Dual-Stream Processing bottleneck and an Efficient Path Aggregation FPN to counteract turbidity-induced degradation, achieving state-of-the-art accuracy with reduced computational cost for smart aquaculture monitoring.

The Gist

Imagine you are trying to take a clear photo of a school of fish swimming in a murky, cloudy pond. It's a nightmare for a camera. The water acts like a dirty window: it blocks certain colors (making everything look blue-green), it scatters light like fog (creating a "veil" over the image), and it blurs the edges of the fish.

Most computer programs trying to "see" these fish today are like a student trying to solve a math problem by brute force. They throw more and more complex tools at the problem, hoping that if they just look harder, they'll figure it out. But this makes the software slow, heavy, and expensive to run on small devices like underwater cameras or drones.

FinSight-Net is a new, smarter approach. Instead of just "trying harder," it understands why the water is messing up the picture and fixes the specific problems one by one. Think of it as a specialized pair of glasses designed just for underwater vision.

Here is how it works, broken down into simple analogies:

1. The "Specialized Team" (MS-DDSP)

Imagine you have a team of four detectives trying to solve a crime in a foggy room.

• The Old Way: You put all four detectives in a single room and tell them to "look at everything." They get confused, and the fog (noise) hides the clues.
• The FinSight-Net Way: You split them into four specialized teams, each with a different job:
  • Team 1 (The Big Picture): Looks for the general shape of the fish, ignoring small details that might be blurry.
  • Team 2 (The Fog Remover): Specifically trained to ignore the "fog" (backscattering noise) caused by floating dirt in the water.
  • Team 3 (The Color Fixer): Knows that red light disappears quickly underwater, so it works extra hard to restore the missing red colors so the fish looks natural.
  • Team 4 (The Detail Keeper): Makes sure the tiny, important bits (like fins or scales) don't get lost in the process.

At the end, a smart manager (an attention mechanism) listens to all four teams and says, "Okay, the fog is thick today, so I'll listen more to Team 2. The colors are off, so I'll trust Team 3 more." This creates a perfect, clear picture of the fish.

2. The "Express Elevator" (EPA-FPN)

In deep learning, information usually travels down a long, winding staircase. As it goes down, it gets heavier and loses details (like a message passed down a line of people that gets garbled). By the time it reaches the bottom, the fine details of the fish are gone.

• The Old Way: The message has to go down every single step, getting slower and losing clarity.
• The FinSight-Net Way: They build an Express Elevator. This allows the "high-definition" details from the top of the building (the raw image) to zip straight down to the bottom, bypassing the slow, noisy staircase. This ensures that even the tiniest edge of a fish fin is preserved, making it much easier to pinpoint exactly where the fish is.

3. Why It Matters

• It's Lighter: Because it's so smart about how it works, it doesn't need a supercomputer to run. It can run on small, battery-powered devices attached to fish farms.
• It's Faster: It processes images in real-time, which is crucial for things like automatic feeding systems. If a fish is hungry, the system needs to know now, not five minutes later.
• It's More Accurate: In tests, it found fish in very cloudy water much better than the current best systems (beating the latest YOLO models by a significant margin).

The Bottom Line

FinSight-Net is like giving a robot fisherman a pair of smart glasses that know exactly how water distorts light. Instead of guessing, it actively cleans up the fog, restores the missing colors, and keeps the fine details sharp. This allows farmers to monitor their fish health and feeding automatically, even in the murkiest ponds, without needing expensive, heavy equipment.

Technical Summary

1. Problem Statement

Underwater fish detection (UFD) is critical for smart aquaculture (e.g., biomass estimation, automated feeding). However, existing deep learning detectors face significant challenges in underwater environments due to intrinsic optical constraints that are often ignored by standard "black-box" models:

• Wavelength-Dependent Absorption: Water absorbs red light faster than blue-green light, causing chromatic distortion and loss of color information.
• Turbidity-Induced Scattering: Suspended particles cause backscattering ($I_{bs}$), creating a "veiling effect" that blurs high-frequency structural details (e.g., fins, edges) and introduces noise.
• Resource Constraints: Many state-of-the-art (SOTA) models rely on heavy attention mechanisms or deep stacking, resulting in high computational overhead unsuitable for edge devices (e.g., AUVs, cage monitors) in aquaculture.
• Localization Failure: Standard Feature Pyramid Networks (FPNs) often lose fine-grained spatial information during deep feature fusion, leading to inaccurate localization of non-rigid fish bodies in turbid water.

2. Methodology: FinSight-Net

The authors propose FinSight-Net, a lightweight, physics-aware detection framework designed to explicitly counteract underwater optical degradation rather than treating it as generic noise. The architecture is built upon a CSPDarknet backbone and integrates two novel components:

A. Multi-Scale Decoupled Dual-Stream Processing (MS-DDSP) Bottleneck

This module addresses frequency-specific information loss by decoupling feature extraction into four parallel heterogeneous branches, each targeting a specific physical degradation:

1. Scale Robustness Branch: Uses staged receptive field expansion to capture multi-scale morphology of non-rigid fish, compensating for global contour blurring.
2. Turbidity Suppression Branch: Focuses on spatial-channel decorrelation to filter out low-value background clutter and pseudo-features caused by suspended particles (backscattering).
3. Absorption Compensation Branch: Employs an inter-channel weighting mechanism to automatically strengthen attenuated spectral channels (e.g., red spectrum), restoring chromatic discriminability.
4. Detail Preservation Branch: Retains a portion of the original feature map without non-linear processing to salvage fine-grained spatial cues (scales, fin edges).

• Mechanism: A soft attention mechanism (Signal-to-Noise Ratio optimization) dynamically weights these branches based on real-time environmental conditions, ensuring the network prioritizes the most robust features.

B. Efficient Path Aggregation FPN (EPA-FPN)

This module acts as a "detail-filling mechanism" to bridge the gap between low-level spatial precision and high-level semantics:

• Long-Range Skip Connections: Instead of standard iterative fusion, EPA-FPN injects high-resolution spatial cues from shallow layers directly into deep semantic nodes. This salvages high-frequency structural information typically filtered out by scattering.
• Path Pruning: Redundant horizontal cross-connections are pruned to reduce computational overhead, resulting in a high-efficiency single-path cross-connection topology.

3. Key Contributions

1. Physics-Aware Framework: FinSight-Net breaks the conventional black-box paradigm by explicitly modeling underwater optical constraints (absorption and scattering) into the network architecture.
2. MS-DDSP Module: A novel bottleneck that achieves frequency-domain feature compensation through parallel decoupling, effectively suppressing backscattering noise while recovering biological structural details.
3. EPA-FPN Module: A lightweight feature aggregation mechanism that enhances semantic-spatial complementarity via long-range skip connections, enabling precise localization with significantly fewer parameters.
4. Efficiency: The network is optimized for resource-constrained platforms, achieving high accuracy with a reduced parameter count compared to standard YOLO variants.

4. Experimental Results

The authors evaluated FinSight-Net on three datasets: DeepFish (natural habitats), AquaFishSet (intensive farming with occlusion), and UW-BlurredFish (a new benchmark simulating extreme turbidity and blur).

• Performance on UW-BlurredFish:
  • Achieved 92.8% mAP, outperforming the SOTA YOLOv11s by 4.8%.
  • Reduced parameters by 29.0% compared to YOLOv11s.
  • Demonstrated superior recall in severe backscattering scenarios compared to heavy models like RT-DETR.
• Generalization:
  • Achieved 94.6% mAP on DeepFish and 93.6% mAP on AquaFishSet, proving robustness across diverse ecological backgrounds and high-density occlusions.
• Ablation Studies:
  • Removing the MS-DDSP or EPA-FPN modules resulted in significant performance drops (e.g., removing the Turbidity Suppression branch dropped mAP by 3.2%).
  • EPA-FPN outperformed PANet and Bi-FPN in both accuracy and inference speed (8.6ms latency).
• Visualization: Feature map analysis confirmed that FinSight-Net significantly improves the Signal-to-Noise Ratio (SNR) and restores high-frequency structural details (fin edges) that are lost in baseline models.

5. Significance

• Scientific Impact: The paper shifts the paradigm of underwater detection from generic noise reduction to physics-informed deep learning, explicitly addressing the optical physics of the medium.
• Practical Application: By offering a lightweight solution (6.7M parameters) that outperforms heavier models, FinSight-Net enables real-time, automated monitoring on edge devices in smart aquaculture, facilitating sustainable fishery management.
• Benchmarking: The introduction of the UW-BlurredFish dataset provides a rigorous stress test for evaluating frequency-domain compensation capabilities in extreme industrial conditions.

In conclusion, FinSight-Net demonstrates that integrating physical constraints into network architecture yields superior robustness and efficiency for underwater vision tasks compared to purely data-driven, black-box optimization approaches.