Improved lightweight YOLOv11 remote sensing small target detection algorithm based on KernelWarehouse

In the field of remote sensing image processing, small object detection has long been confronted with numerous technical challenges. To address critical issues in current dynamic convolution methods for remote sensing small object detection, such as low parameter utilization and the mismatch between model capacity and computational resource requirements, this study proposes an improved YOLOv11 model based on an optimized KernelWarehouse architecture. The innovative solution systematically enhances the dynamic convolution mechanism through three key optimizations: implementing a kernel partitioning strategy to divide convolution kernels into multiple functional regions, enabling each partition to focus on processing features at different scales; constructing a shared warehouse architecture that facilitates cross-layer and cross-scale weight reuse through a parameter sharing pool; and introducing a contrastive attention function module to effectively enhance the model's capability in capturing small object features. Notably, this research incorporates channel pruning technology by establishing an importance evaluation mechanism to intelligently screen and streamline redundant network structures, significantly reducing computational complexity while maintaining detection performance. To validate the model's effectiveness, comparative experiments were conducted on the SIMD benchmark remote sensing dataset. Results demonstrate substantial improvements across multiple key metrics: a 1.0% increase in mean average precision (mAP), a 29.3% reduction in model parameters, and a 12.8% enhancement in small object detection accuracy. These improvements substantiate the innovative value of the proposed solution in addressing the parameter efficiency bottleneck of dynamic convolution. This study not only provides a novel approach for achieving synergistic optimization of accuracy and efficiency in remote sensing small object detection tasks, but also offers a parameter sharing framework and model lightweighting methodology that can be extended to other computer vision applications, demonstrating significant reference value for both theoretical algorithm research and practical engineering implementations.