Lml-yolo: Composite enhanced YOLO object detection model for underwater complex environments

Abstract Underwater object detection encounters technical challenges such as complex background interference, small target scales, random distribution directions, and diverse biomorphology. To address these critical issues, this study proposes LML-YOLO, an enhanced lightweight model based on YOLO11n, which integrates the Lightweight Receptive-Field Attention Convolutional (LRFAConv) module and Multidimensional Directional Collaborative Attention (MDCA). LRFAConv dynamically optimizes convolutional kernel parameters through a receptive-field attention mechanism, effectively enhancing target feature extraction while suppressing background interference. MDCA extends multidimensional spatial collaborative attention by incorporating diagonal and anti-angle directional attention, improving multi-angle feature perception. Additionally, the Neck module uses linear deformable attention convolution (LDAConv), combining deformable convolution with a perceptual field attention mechanism to better adapt to diverse biological morphologies. Experimental results demonstrate that the proposed LML-YOLO achieves mAP scores of 65.2%, 29.8%, and 55% on DUO, UDD, and UODD datasets, respectively, which outperform the original model by 2.1%, 1.4%, and 1.2%, while reducing parameters and GFLOPs by 6.9% and 4.7%. These advancements establish a novel method in efficient underwater object detection, particularly beneficial for marine biological resource monitoring and biodiversity conservation applications.