YOLO-PKFF: Remote Sensing Object Detection Enhanced With Poly Kernel Inception and Attentional Cross-Level Feature Fusion

With the rapid development of satellite and deep learning technologies, remote sensing object detection has emerged as a research hotspot. However, recent research mainly focuses on representing bounding boxes and feature extraction to improve detection accuracy without considering the inherent characteristics of remote sensing scenes. A novel remote sensing object detector based on YOLO11 with the poly kernel inception and an attentional cross-level feature fusion (YOLO-PKFF) is proposed to bridge this gap. First, the poly kernel inception network (PKINet) is introduced as the backbone network to effectively capture local and global contextual information. Second, the attentional cross-level feature fusion (ACFF) module is used to selectively integrate low-level texture features with high-level semantic features. Finally, an enhanced inception module is integrated into the C3k2 module, improving the detection of striped objects. Experimental results demonstrate that the proposed YOLO-PKFF achieves a 3.8% improvement in mean average precision (mAP) on the DIOR optical remote sensing dataset compared with the baseline. Furthermore, YOLO-PKFF achieves mAP improvements of 6.2% and 1.2% on two additional publicly available optical remote sensing datasets, respectively.