MDCT: Multi-Kernel Dilated Convolution and Transformer for One-Stage Object Detection of Remote Sensing Images

Deep learning (DL)-based object detection algorithms have gained impressive achievements in natural images and gradually matured recent years. However, compared with images, remote sensing are faced severe challenges due to the complex backgrounds difficult of small objects dense scenes. To address these problems, a novel one-stage model named MDCT is proposed based on multi-kernel dilated convolution (MDC) block transformer block. Firstly, new feature enhancement module, MDC block, developed enhance objects’ ontology adjacent spatial features. Secondly, we integrate into neck network order prevent loss information Finally, depthwise separable introduced each reduce computational cost. We conduct experiments three datasets: DIOR, DOTA, NWPU VHR-10. Compared YOLOv5, our improves accuracy by 2.3%, 0.9%, 2.9% VHR-10 datasets, respectively.