Remote Sensing Object Detection Method Based on Three-Dimensional Weight-Guided Selection Kernel

In recent years, research on remote sensing object detection has primarily focused on oriented box representation and optimization, while overlooking the importance of studying feature extraction networks, leading to inaccurate identification of some targets. The conventional feature extraction networks employ fixed-size convolutional kernels, which fail to capture different levels of visual features for targets of varying sizes. In this paper, we propose a mechanism called the 3D-Weight Guide Selective Kernel (3D-WGSK) to jointly utilize spatial and channel information of input features to derive three-dimensional weights. These weights are then employed for convolutional kernel selection, enabling adaptive receptive field acquisition for capturing features of targets with different sizes. We devise a network architecture named the 3D-WGSK Net for feature extraction in remote sensing object detection. It possesses the capability of multi-scale information fusion and adaptive kernel adjustment, effectively extracting crucial features of targets with varying sizes. Experimental results on the publicly available DOTA-v1.5 dataset demonstrate the superior detection performance of our method.