[Paper Review] YOLO series target detection algorithms for underwater environments

Chenjie Zhang, Pengcheng Jiao|arXiv (Cornell University)|Sep 7, 2023

Water Quality Monitoring TechnologiesEnvironmental Science3 citations

TL;DR

This paper proposes YOLO-UC and YOLO-UH, improved YOLO-based models for underwater target detection in structural health monitoring and toxic organism detection. By integrating data augmentation, Laplace sharpening, and small-object optimization, the models enhance detection accuracy under low-visibility conditions, achieving improved robustness despite challenges like turbidity, blur, and limited recognition distance.

ABSTRACT

You Only Look Once (YOLO) algorithm is a representative target detection algorithm emerging in 2016, which is known for its balance of computing speed and accuracy, and now plays an important role in various fields of human production and life. However, there are still many limitations in the application of YOLO algorithm in underwater environments due to problems such as dim light and turbid water. With limited land area resources, the ocean must have great potential for future human development. In this paper, starting from the actual needs of marine engineering applications, taking underwater structural health monitoring (SHM) and underwater biological detection as examples, we propose improved methods for the application of underwater YOLO algorithms, and point out the problems that still exist.

Motivation & Objective

• Address the lack of robust target detection in underwater environments due to poor visibility, turbidity, and image degradation.
• Develop YOLO-UC for detecting underwater concrete damage in marine structures, improving detection accuracy under challenging imaging conditions.
• Create YOLO-UH for identifying small, potentially toxic marine organisms, focusing on enhancing detection of small targets in real-world underwater settings.
• Investigate hardware and software limitations affecting YOLO deployment in underwater robotics and autonomous systems.
• Identify key challenges such as camera quality, computational constraints, communication limitations, and domain shift from terrestrial to underwater domains.

Proposed method

• Applied photometric and geometric data augmentation to improve model robustness to lighting and perspective variations in underwater scenes.
• Integrated mosaic and mix augmentation techniques to enhance detection performance on complex backgrounds and small objects.
• Replaced standard IoU loss with Generalized IoU (GIoU) loss to improve bounding box regression, especially in non-overlapping prediction scenarios.
• Applied Laplace sharpening to enhance image clarity and reduce blur in underwater images, improving feature extraction.
• Enhanced small-object detection in YOLO-UH using specialized neck modules and data augmentation tailored for tiny targets.
• Proposed an early warning mechanism to reduce missed detections by prioritizing and refining output bounding boxes.

Experimental results

Research questions

• RQ1 How can YOLO-based models be effectively adapted to underwater environments with low visibility, turbidity, and image degradation?
• RQ2 What data augmentation and image preprocessing techniques significantly improve detection accuracy in underwater target detection tasks?
• RQ3 To what extent can small-object detection performance be enhanced in underwater settings using model architecture modifications?
• RQ4 What are the primary hardware and software limitations affecting real-time YOLO deployment in underwater robotic systems?
• RQ5 How does domain shift from terrestrial to underwater environments impact YOLO model generalization, and what adaptation strategies are effective?

Key findings

• The YOLO-UC model achieved improved detection accuracy on underwater concrete damage using GIoU loss and data augmentation, particularly under variable lighting and perspective conditions.
• YOLO-UH demonstrated enhanced small-object detection capability through specialized neck modules and data augmentation, outperforming baseline YOLOv5 in detecting tiny marine organisms.
• Recognition accuracy dropped significantly beyond 1.5 meters due to camera focus failure and image blur, indicating a hard limit on effective detection range.
• Image preprocessing techniques such as Laplace sharpening and contrast enhancement significantly improved feature visibility and detection reliability.
• Field tests revealed that sediment concentration and distance strongly affect detection performance, with accuracy declining under high turbidity and long distances.
• Hardware constraints—especially camera resolution, autofocus, and chip power—were identified as major bottlenecks, limiting overall system performance despite software improvements.