Research on Object Detection of High Resolution Remote Sensing Image based on Improved YOLOV4 Algorithm

Abstract Object detection is one of the fundamental tasks in computer vision. Although excellent progress has been made, there still exist challenges for objects with dense distribution, fuzzy feature, and small size. This paper presents a detector for object detection of small size in High-Resolution remote sensing images, namely SF-YOLOV4. The proposed SF-YOLOV4 exploits information for shallow layers and contextual information along with spatial attention to address the above challenges. Specifically, a shallow semantic information extraction network(SFN) is designed which introduces low-level semantic information into the backbone, to alleviate the loss of the small object features. Meanwhile, we replace the original Nick with multi-scale context feature pyramid (MSC-FPN) to improve the utilization of lower layers information and integrate the context information, we also add spatial attention module to find attention region at different scales. Experiments on two remote sensings public datasets DIOR and RSOD show the good detection performance of SF-YOLOV4.