Abstract: Object detection in remote sensing images presents numerous challenges, such as significant variations in object scale and pronounced differences in aspect ratios. Current mainstream approaches typically employ large convolutional kernels or multi-scale convolutional architectures to address these challenges. However, large convolutional kernels tend to introduce substantial background noise, while the use of kernels with varying scales may lead to feature information redundancy. To tackle these issues, we propose a robust and efficient rotation-aware object detection method–MKS-YOLO. MKS-YOLO explicitly decouples high- and low-informative features through an information-aware threshold gating mechanism, effectively suppressing background noise interference. Furthermore, to reduce noise and enhance the model’s ability to perceive objects at multiple scales, we design a Multi-Kernel Dynamic Selection Module (MKSM). This module integrates multiscale convolutional kernels and introduces a spatial attention mechanism to achieve dynamic weighted feature fusion. In addition, we incorporate a Contextual Large Kernel Attention (CLKA) mechanism to enhance feature representation capabilities for objects with high aspect ratios. Moreover, during the feature extraction stage, we introduce an Attention-based Internal Feature Interaction module (AIFI), which further strengthens the correlations among deep semantic features. To validate the effectiveness of the proposed method, we conduct extensive experiments on two widely used remote sensing image datasets: DOTA-v1.0 and DIOR-R. Experimental results demonstrate that MKS-YOLO achieves state-of-the-art detection accuracy, with Mean Average Precision (mAP) reaching 77.71% and 82.50%, respectively, fully validating the superior performance of the proposed method.

Key words: Object detection')">Object detection, Detectors, Remote sensing, Dynamic selection network