ODConv-GFPN-LSPCD: A Multi-Module Enhanced YOLOv11n for Marine Organism Detection
This paper proposes ODConv-GFPN-LSPCD, a multi-module enhanced YOLOv11n architecture that integrates Omni-dimensional Dynamic Convolution, a Giraffe Feature Pyramid Network, and a Lightweight Shared Convolutional Detection Head to achieve superior accuracy and real-time efficiency for detecting marine organisms in challenging underwater environments.
Original paper licensed under CC BY 4.0 (https://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
The ocean is a vast, shifting world where light behaves differently than it does on land. Sunlight scatters and fades as it penetrates the water, creating scenes that are often dim, murky, and filled with obstacles. For scientists and conservationists, understanding what lives beneath the surface is critical for protecting ecosystems and managing fisheries, but watching these environments is incredibly difficult. Humans cannot stay underwater for long, and traditional methods like dragging nets can damage the very habitats they are trying to study. This has led to a reliance on cameras and robots, yet teaching computers to "see" underwater is a unique challenge. Standard vision systems, trained on clear images of cars or people on dry land, often fail when faced with the distortion, low light, and small, moving creatures of the sea. They miss tiny animals, lose track of creatures that bend and twist, and struggle when visibility is poor.
To solve this, researchers have turned to artificial intelligence, specifically a type of computer vision system known as a detector. These systems scan an image and draw boxes around objects, identifying what they are and where they are located. While modern detectors are powerful, the ones designed for speed and efficiency often lack the precision needed for the complex underwater world. A recent study from Zhengzhou University and Sichuan Agricultural University addresses this gap by refining a fast, lightweight detection system specifically for marine life. The team did not invent a new system from scratch; instead, they took an existing, efficient model and surgically improved three specific parts of its internal logic. Their goal was to create a tool that is both fast enough to run on small, battery-powered robots and accurate enough to spot a small sea star or a camouflaged fish in a cluttered environment.
The researchers began with a baseline model called YOLOv11n, which is already known for being quick and compact. However, they identified three specific weaknesses that made it struggle underwater. First, the model's initial layers, which reduce the size of an image to process it, used a rigid, static method that could not adapt to the strange shapes of marine animals. Second, the part of the system that combines information from different distances—helping the computer see both tiny details and large shapes together—was too fixed, causing it to lose track of small or hidden creatures. Third, the final layer that makes the actual guess about what an object is carried too much unnecessary weight, making the system prone to errors when trained on limited data.
To fix these issues, the team introduced three complementary upgrades. The first change replaced the rigid initial layers with a dynamic system that can adjust its focus. Imagine a camera lens that can instantly reshape its glass to match the curve of a fish's body rather than forcing the fish to fit a standard shape. This allowed the system to better capture the details of animals that twist, turn, or have flexible bodies. The second upgrade reorganized how the system shares information between different layers of processing. Instead of a one-way street where information flows in a single direction, they created a network where details could circulate and reinforce each other. This ensured that the tiny features of a small shrimp or a distant jellyfish were not lost as the image was processed. The third and most impactful change was a streamlined final layer. By removing redundant parts and sharing resources between different detection tasks, the researchers significantly reduced the amount of memory the system required without sacrificing its ability to be accurate.
When the team tested these improvements on a dataset of nearly ten thousand underwater images containing eleven different types of marine life, the results were clear. The modified system, which they named ODConv-GFPN-LSPCD, outperformed the original model and several other leading detectors. In a direct comparison, the new system achieved a higher accuracy rate in identifying and locating objects, reaching a score of 47.20 percent on their specific test, while using fewer computing resources than the baseline. It successfully identified small and partially hidden creatures that other models missed, such as clownfish hiding in coral or bivalves on the seafloor. The system also maintained a high speed, processing images quickly enough to be useful for real-time applications on underwater robots.
The study highlights that the most significant gains came from the streamlined final layer, which proved that making a system smaller and less cluttered could actually make it smarter by reducing confusion. While the dynamic adjustments for shape and the improved information sharing helped, the reduction in unnecessary complexity was the key driver of the improved performance. The researchers noted that the system still faces challenges in extremely dark water or when certain rare species are underrepresented in the training data, but the overall balance of speed and accuracy represents a major step forward. By creating a tool that is both lightweight and precise, this work offers a practical path toward deploying autonomous underwater vehicles that can monitor marine biodiversity continuously, helping to protect the ocean's hidden life without the need for constant human intervention.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.