Comprehensive Attention Mechanism: A New Attention Mechanism Integrating Global Information and Local Details

This research presents a novel Comprehensive Attention Mechanism (CAM) aimed at addressing the shortcomings of current attention mechanisms in effectively integrating local details with global semantics. CAM is engineered to overcome specific challenges, such as the detection of small objects and the preservation of fine-grained features in medical imaging. The key contributions of this study include: (1) Deformable Convolution and Dynamic Alignment: This component employs deformable convolution to ensure accurate alignment of multiscale features, thereby reducing cross-scale feature misalignment; (2) Dynamic Weight Generation: Through an entropy-based dynamic gating network, this mechanism produces task-specific fusion weights, facilitating fine-grained regulation from the pixel level to the scene level; (3) Sparse Attention and Lightweight Computation: CAM incorporates a sparse attention strategy to significantly cut down on redundant computations, alongside lightweight local convolution and global interaction modules, achieving up to a 70 % reduction in computational cost while maintaining effective multiscale feature integration. Experimental results indicate that CAM enhances the average precision (AP) by 12 % on the DOTAv1 and VisDrone2019 datasets, providing a robust solution for complex visual tasks.