💻 computer science

A Marine Debris Detection Framework for Ocean Robots via Self-Attention Enhancement and Feature Interaction Optimization

This paper proposes YOLO-MD, an enhanced YOLO-based framework featuring a Dual-Branch Convolutional Enhanced Self-Attention module, a lightweight shift-based operation, and SFG-Loss to achieve state-of-the-art marine debris detection performance on low-quality images and verify its effectiveness through real-world robotic edge deployment.

Original authors: Yuyang Li, Jiashu Han, Yinyi Lai, Wenbin Kang, Zenghui Liu

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yuyang Li, Jiashu Han, Yinyi Lai, Wenbin Kang, Zenghui Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the ocean as a giant, messy attic filled with hidden treasures and trash. Our goal is to send out a robot (an "ocean robot") to find and pick up the trash (marine debris). But there's a catch: the attic is dark, foggy, and the trash is often tiny, blurry, or hiding behind piles of seaweed. Trying to find a lost earring in a foggy room is hard; finding a plastic bottle in a blurry underwater photo is even harder for a computer.

This paper introduces a new "super-eye" for these robots called YOLO-MD. Think of it as upgrading a standard pair of glasses to a high-tech, smart-lens system specifically designed for messy, underwater conditions.

Here is how the authors built this smarter system, using three main tricks:

1. The "Dual-Branch" Glasses (DB-CASA)

The Problem: Standard computer vision often gets distracted by the big picture (the whole ocean) and misses the tiny details (a small piece of plastic). It's like trying to read a tiny label on a bottle while looking at the whole ocean from a helicopter.
The Solution: The authors created a special module called DB-CASA. Imagine this as a pair of glasses with two lenses working at once:

Lens A (Spatial): Focuses on the shape and edges of objects, like tracing the outline of a bottle.
Lens B (Channel): Focuses on the colors and textures, like noticing the specific shade of blue on a plastic bag.
Instead of just looking at the whole scene, this module forces the robot to look at the shape and the color separately and then combine them. This helps the robot spot small, blurry items that usually get lost in the "noise" of the water.

2. The "Shifting Puzzle" (Feature Shift Fusion)

The Problem: Sometimes, the computer sees a tiny object, but the details are too spread out to be useful. It's like trying to solve a puzzle where the pieces are slightly misaligned.
The Solution: They introduced a Feature Shift Fusion Module. Imagine you have a grid of puzzle pieces. Instead of just gluing them together, this module gently shifts some pieces left, right, up, or down, and then snaps them back together.

This "shifting" happens without adding any heavy new parts (it's "parameter-free," meaning it doesn't make the robot slower or heavier).
By nudging the information around, the robot can better connect the dots to see exactly where a small piece of trash is, even if it's tiny or partially hidden.

3. The "Fair Teacher" (SFG-Loss)

The Problem: When teaching a robot to find trash, some pictures are easy (a big, clear bottle), and some are hard (a tiny, blurry wrapper). If the robot gets too good at the easy ones, it stops trying to learn the hard ones. This is called "class imbalance."
The Solution: They created a new scoring system called SFG-Loss. Think of this as a strict but fair teacher.

If the robot gets an easy question right, the teacher gives a small "good job" (low weight).
If the robot struggles with a hard question (like a blurry target), the teacher says, "Pay attention here!" and gives that question extra importance (high weight).
This ensures the robot keeps practicing on the difficult, tricky cases until it masters them, rather than just coasting on the easy wins.

The Results: Does it work?

The authors tested this new "super-eye" on a dataset called UODM (a collection of 5,000+ underwater photos of trash).

The Score: YOLO-MD scored higher than almost every other method tested, including older "two-stage" detectives and newer "Transformer" models. It achieved a precision of 0.875 (meaning when it says it found trash, it's usually right) and an F1-score of 0.822.
The Real-World Test: They didn't just test it on a computer. They put it on a real unmanned boat in a campus lake. Even though the water was murky, the light was dim, and the images were blurry, the robot successfully found and located trash in real-time without needing to call a supercomputer back at the lab for help.

Summary

In short, the paper claims that by giving the robot better focus (Dual-Branch), smarter alignment (Shifting), and fairer training (SFG-Loss), we can make ocean robots much better at finding trash in the messy, blurry, real world. This helps protect the ocean without slowing the robots down.

Technical Summary: YOLO-MD for Marine Debris Detection

Problem Statement
Accurate detection and localization of marine debris are critical for ecological protection and pollution control, particularly for ocean robots. However, detection performance is frequently degraded by the challenging nature of underwater imagery, which often suffers from blur, complex backgrounds, and the presence of small, densely distributed targets. Traditional image enhancement techniques (e.g., deblurring, super-resolution) introduce prohibitive computational overhead, making them unsuitable for real-time applications. Furthermore, while existing object detectors like YOLO offer a balance of speed and accuracy, they struggle in blurred scenes and with small-object detection. Transformer-based models provide strong global modeling but often lack sensitivity to local fine-grained features and incur high computational costs, limiting their deployment in resource-constrained robotic scenarios. Additionally, class imbalance and optimization instability in complex marine environments remain significant hurdles.

Methodology
To address these challenges, the authors propose YOLO-MD, an enhanced YOLO-based detection framework designed to improve feature representation, local interaction, and training stability. The architecture integrates three primary innovations:

Dual-Branch Convolutional Enhanced Self-Attention (DB-CASA):
Replacing the standard C2PSA module in the YOLO backbone, DB-CASA is designed to strengthen spatial–channel interactions without the high cost of global attention. It decomposes input features into two complementary branches:
- A Spatial Interaction Branch that uses depth-wise convolution and a $1 \times 1$ convolution to generate a spatial attention map, enhancing boundary and texture features of small targets.
- A Channel Interaction Branch inspired by SENet, which combines global average pooling with convolution to capture inter-channel dependencies without reducing channel counts.
- The module employs an additive fusion strategy for the Query and Key outputs of these branches, avoiding expensive similarity matrix computations while preserving fine-grained features often suppressed by global aggregation.
Feature Shift Fusion Module (FSFM):
To enhance local feature interaction with minimal overhead, FSFM introduces a parameter-free shift-based operation. The input feature map is divided into sub-groups along the channel dimension, which are then shifted in four directions (horizontal and vertical). This cross-dimensional feature exchange allows the model to capture multi-orientation local context. Integrated within the C3k2 structure with residual connections, FSFM refines localization precision without adding learnable parameters.
SFG-Loss (Slide-Focaler-GIoU Loss):
To mitigate class imbalance and optimization instability, the authors propose a composite loss function. It combines:
- Generalized IoU (GIoU) as the base regression loss for stable gradients.
- A dynamic weighting mechanism (inspired by Slide Loss) that adaptively emphasizes hard samples while suppressing easy ones.
- Focaler-IoU constraints to truncate excessively easy or hard outliers, preventing training instability.
  This unified framework ensures robust convergence in complex detection scenarios.

Key Contributions
The paper summarizes its contributions as follows:

The design of the DB-CASA module, which utilizes additive similarity and parallel spatial-channel interactions to preserve fine-grained features of small targets in low-quality images.
The development of the Feature Shift Fusion Module (FSFM), which employs a cross-dimensional feature exchange strategy via grouped spatial shifts to enhance small-object detection precision without increasing parameter count.
The proposal of SFG-Loss, a dynamic sample reweighting mechanism designed to address class imbalance and optimization instability.

Experimental Results
Experiments were conducted on the UODM dataset (5,125 images, 15 categories) using an NVIDIA RTX 3070 GPU. The proposed YOLO-MD achieved the following performance metrics:

Precision: 0.875
F1-score: 0.822
mAP50: 0.849

YOLO-MD outperformed a wide range of state-of-the-art detectors, including RT-DETR-l, Faster R-CNN, SSD, and various YOLO variants (v5, v8, v10, v11, v12, v13, and YOLO26). Notably, compared to the baseline YOLOv11, YOLO-MD improved mAP50 by 2.8% and Precision by 4.0%, while maintaining a comparable parameter count (2.78M). Ablation studies confirmed that the integration of all three modules (DB-CASA, FSFM, and SFG-Loss) yields optimal performance, resolving trade-offs between global attention and local shifting.

Significance and Real-World Validation
The paper claims that YOLO-MD offers a favorable trade-off between detection accuracy and computational efficiency, making it suitable for real-time, resource-constrained ocean robots. Beyond benchmark performance, the method's practical applicability was verified through real-world edge deployment. The system was tested on an unmanned surface vehicle (USV) equipped with an underwater camera and embedded edge computing device in a campus artificial lake. Despite challenging field conditions—including low illumination, high turbidity, and image blur—YOLO-MD successfully localized and recognized submerged debris in real-time. The authors conclude that the framework effectively addresses the specific degradation issues of marine environments, providing a robust solution for ecological protection and robotic perception.