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

This paper proposes a lightweight underwater object detection framework based on YOLOv10 that integrates a Multi-Stage Adaptive Enhancement module, a Dual-Pooling Sequential Attention mechanism, and a Focal Generalized IoU loss to significantly improve accuracy and robustness on benchmark datasets while maintaining a compact model size suitable for resource-constrained environments.

The Gist

Imagine you are trying to spot a rare, colorful fish swimming in a murky, deep ocean. The water is cloudy, the light is dim, and everything looks greenish-blue or blurry. Now, imagine you are a robot (an underwater drone) trying to find that fish using a camera. It's incredibly hard because the camera sees a mess of colors and shadows, not a clear picture.

This paper is about teaching that robot a new, super-smart way to see clearly in that messy water, without needing a giant, expensive computer to do the thinking.

Here is the breakdown of their solution, using simple analogies:

1. The Problem: The "Murky Water" Effect

Underwater cameras struggle because water eats away at light and scatters it. It's like trying to read a book through a dirty, foggy window.

• The Issue: Standard AI models (the "brains" of the robot) were trained on clear, sunny land photos. When they look underwater, they get confused. They can't tell the difference between a fish and a rock because the colors are wrong and the edges are fuzzy.
• The Consequence: The robot misses the fish or mistakes a bubble for a fish.

2. The Solution: A Three-Step "Super-Vision" Kit

The authors built a new system based on a popular AI model called YOLOv10 (which is like a very fast, efficient detective). They gave this detective three special tools to handle the underwater mess:

Tool A: The "Digital Photo Editor" (Multi-Stage Adaptive Enhancement)

Before the detective even looks at the fish, they first clean up the photo.

• The Analogy: Imagine you have a photo that looks too blue and dark. You use a filter to add red back in, brighten the shadows, and sharpen the edges.
• What they did: They created a step-by-step process that automatically fixes the color (removes the blue tint), boosts the contrast (makes dark things lighter), and removes the "haze" (like fog). Crucially, this is a fixed rulebook, not a learning process, so it happens instantly without slowing the robot down.

Tool B: The "Binoculars with a Spotlight" (Dual-Pooling Sequential Attention)

Once the photo is cleaned, the detective needs to know where to look.

• The Analogy: Imagine you are in a crowded room. Instead of staring at the whole room, you put on binoculars that zoom in on specific spots (spatial attention) and then filter out the noise so you only see the person you are looking for (channel attention).
• What they did: They added a "Dual-Pooling Sequential Attention" mechanism. It acts like a spotlight that ignores the boring background (sand, bubbles, seaweed) and focuses intensely on the small, important objects (fish, turtles). It helps the robot see tiny details that usually get lost in the noise.

Tool C: The "Strict Coach" (FGIoU Loss)

When the robot makes a guess ("That's a fish!"), it needs to be graded on how good the guess was.

• The Analogy: Imagine a coach grading a student. If the student says "It's a fish" but draws the box around the fish too loosely, the coach says, "No, that's not precise enough." If the student misses a fish entirely, the coach says, "You missed it!"
• What they did: They created a new scoring system (a "Loss Function") that punishes the robot if it's not precise with the box around the fish or if it gets confused about whether an object is actually there. It forces the robot to be both accurate and confident.

3. The Result: Fast, Small, and Super Accurate

The best part? They didn't need a supercomputer to run this.

• The "Lightweight" Magic: Usually, making AI smarter makes it heavier and slower. But this team managed to make the AI smarter and keep it tiny (only 2.8 million "parameters," which is like the size of a small app on your phone).
• The Score: When they tested it on real underwater datasets (RUOD and DUO), their new system was 6% to 7% more accurate than the standard version.
  • Think of it this way: If the old robot found 82 out of 100 fish, the new robot finds 89 out of 100. That's a huge difference when you are looking for rare species or navigating safely.

Why Does This Matter?

This technology is perfect for Autonomous Underwater Vehicles (AUVs)—robots that explore the ocean without a human pilot.

• These robots have limited battery and small computers.
• They can't carry a massive supercomputer.
• This new method allows them to see clearly, find objects quickly, and make decisions in real-time, even in the darkest, murkiest parts of the ocean.

In a nutshell: The authors took a standard AI detective, gave it a photo editor to clean the view, a spotlight to focus on the target, and a strict coach to improve its accuracy, all while keeping the detective small enough to fit in a backpack.

Technical Summary

1. Problem Statement

Underwater object detection (UOD) is critical for marine surveillance and autonomous systems but faces severe challenges due to the underwater optical environment. Key issues include:

• Visual Degradation: Wavelength-dependent absorption, scattering, and non-uniform illumination cause color distortion (cyan bias), low contrast, and boundary blurring.
• Model Limitations: Standard detectors trained on terrestrial data fail underwater due to poor early-stage feature extraction.
• Resource Constraints: Existing high-accuracy solutions often rely on computationally heavy enhancement or attention modules, making them unsuitable for real-time deployment on resource-constrained platforms like Autonomous Underwater Vehicles (AUVs) and Remotely Operated Vehicles (ROVs).
• Imbalance: Detection models struggle with class imbalance and localization uncertainty in complex underwater backgrounds.

2. Methodology

The authors propose a lightweight, robust framework based on YOLOv10n, integrating three core components to address the aforementioned challenges:

A. Multi-Stage Adaptive Enhancement (MAE-UVP)

A deterministic preprocessing pipeline designed to restore image quality without learnable parameters (ensuring low overhead and reproducibility). It consists of four sequential phases:

1. Adaptive Color Correction: Compensates for cyan bias via channel-wise scaling to recover attenuated red components.
2. Luminance Contrast Enhancement: Converts images to CIELAB space and applies Contrast Limited Adaptive Histogram Equalization (CLAHE) solely to the luminance channel to boost contrast without color distortion.
3. Soft-Guided Dehazing (SGD): Uses a Gaussian-guided prior to attenuate forward scattering haze while preserving edge clarity and preventing halo artifacts.
4. Edge Preserving Refinement: Applies edge-aware filtering to enhance object boundaries and reduce noise in homogeneous regions.

B. Dual-Pooling Sequential Attention (DPSA)

A lightweight attention mechanism embedded into the backbone, specifically replacing the standard Spatial Pyramid Pooling Fast (SPPF) layer.

• Placement: Applied post-concatenation of multi-scale features to refine representations before the detection head.
• Mechanism: It sequentially applies:
  1. Channel Attention: Uses dual adaptive pooling and a shared 2-layer MLP (reduction ratio 16) to generate channel weights.
  2. Spatial Attention: Computes mean and max statistics across channels, concatenates them, and applies a fixed $7\times7$ convolution.
• Goal: To strengthen small object features and suppress complex underwater background noise without altering the backbone topology.

C. Focal Generalized IoU Objectness Loss (FGIoU)

A hybrid loss function designed to jointly optimize localization, classification, and objectness. It combines three components:

1. Generalized IoU (GIoU) Loss: Improves bounding box regression by penalizing both insufficient overlap and spatial separation.
2. Focal Loss: Addresses foreground-background class imbalance by focusing training on hard examples.
3. Objectness Focal Loss: Enhances confidence calibration using focal-weighted binary cross-entropy.

• Formulation: $L_{FGIoU} = 7.5 \cdot L_{GIoU} + 0.5 \cdot L_{Focal} + 1.0 \cdot L_{ObjFocal}$

3. Key Contributions

1. Deterministic Preprocessing: Introduction of the MAE-UVP module to correct color and contrast issues without adding trainable parameters or inference latency.
2. Novel Attention Module: Development of the DPSA mechanism, which efficiently refines multi-scale features in the backbone to handle small objects and noisy backgrounds.
3. Hybrid Loss Function: Creation of FGIoU, which simultaneously optimizes localization accuracy and objectness calibration under class imbalance.
4. Lightweight Architecture: The final model (DPSA FGIoU YOLOv10n) maintains a compact size of only 2.8M parameters, making it ideal for embedded deployment.

4. Experimental Results

The framework was evaluated on two standard benchmarks: RUOD (10 categories) and DUO (4 categories).

• Performance Gains (vs. Baseline YOLOv10n):
  • RUOD Dataset: Achieved 88.9% mAP@0.5 (a +6.7% improvement over the baseline of 82.2%).
  • DUO Dataset: Achieved 88.0% mAP@0.5 (a +6.2% improvement over the baseline of 81.8%).
• Ablation Study:
  • Adding DPSA alone improved mAP by ~6.1% (RUOD).
  • Adding FGIoU alone improved mAP by ~5.8% (RUOD).
  • The combination yielded the best results, demonstrating complementary benefits.
• Comparison with SOTA:
  • Outperformed other YOLO variants (YOLOv8n/s/m, YOLOv9t, YOLOv10s, YOLOv11n) and specialized underwater models (Dynamic YOLO, CEH-YOLO).
  • Notably, the proposed model achieved higher accuracy than larger models (e.g., YOLOv8m with 25.8M params) while using significantly fewer parameters (2.8M).
• Efficiency:
  • Inference time: 2.1 ms per image (~476 FPS) at 640×640 resolution.
  • Preserves real-time performance suitable for embedded systems.

5. Significance

This paper presents a highly effective solution for the "accuracy vs. efficiency" trade-off in underwater computer vision. By decoupling image enhancement (via deterministic preprocessing) from feature learning and optimizing the loss landscape, the authors achieved state-of-the-art performance on standard benchmarks while maintaining a lightweight architecture. The proposed framework is particularly significant for real-world deployment on AUVs and ROVs where computational power, energy, and memory are limited, yet high-precision detection is required for tasks like ecological monitoring and resource management.