UM-YOLOv10: Underwater Object Detection Algorithm for Marine Environment Based on YOLOv10 Model
Rengui Mai, Wang Ji · Fishes · 2025
In order to address the challenges of a low detection accuracy, missed detections, and false detections in marine precious biological target detection within complex marine environments, this paper presents a novel residual attention module called R-AM. This module is integrated into the backbone network of the YOLOv10 model to improve the model’s focus on the detailed features of biological targets during feature extraction. Additionally, the introduction of a bidirectional feature pyramid with adaptive feature fusion in the neck network enhances the integration of semantic information from deep layers, and localization cues from shallow layers improve the model’s ability to distinguish targets from their environments. The experimental data showed that the improved YOLOv10 model achieved 92.89% at [email protected], increasing by 1.31% compared to the original YOLOv10 model. Additionally, the [email protected]:0.95 was 77.13%, indicating a 3.71% improvement over the original YOLOv10 model. When compared to the Faster R-CNN, SSD, RetinaNet, YOLOv6, and YOLOv7 models, the enhanced model exhibited increases of 1.5%, 1.7%, 4.06%, 4.7%, and 1.42% in [email protected], respectively. This demonstrates a high detection accuracy and robust stability in complex seabed environments, providing valuable technical support for the scientific management of marine resources in underwater ranches.