CEH-YOLO: A composite enhanced YOLO-based model for underwater object detection

Advances in underwater recording and processing systems have highlighted the need for automated methods dedicated to the accurate detection and tracking of small underwater objects in imagery. However, the unique characteristics of underwater optical images, including low contrast, color variations, and the presence of small objects, pose significant challenges. This paper presents CEH-YOLO, a variant of YOLOv8, incorporating a high-order deformable attention (HDA) module to enhance spatial feature extraction and interaction by prioritizing key areas within the model. Additionally, the enhanced spatial pyramid pooling-fast (ESPPF) module is integrated to enhance the extraction of object attributes, such as color and texture, which is particularly beneficial in scenarios with small or overlapping objects. The customized composite detection (CD) module further improves the accuracy and inclusivity of object detection. Moreover, the model uses the WIoU v3 technique for bounding box loss calculations, effectively addressing regression challenges related to bounding boxes under standard and extreme conditions. The experimental results show the model's exceptional performance, achieving mean average precisions of 88.4% and 87.7% on the DUO and UTDAC2020 datasets, respectively. Notably, the model operates at a rapid detection speed of 156 FPS, fulfilling critical real-time detection needs. With a concise model size of 4.4 M and a moderate computational complexity of 11.6 GFLOPs, it is highly suitable for integration into underwater detection systems.