Remote Sensing Target Detection Based on Improved YoloX

Abstract

The identification and positioning of remotely sensed targets hold significant potential in various fields. Nonetheless, the detection and localisation of the algorithm face challenges due to the divergence of size, shape and spatial position distribution of the objects captured in aerial photography. To address this issue, this study commences with YOLOX and implements the GCNet network along with the suggested feature pre-fusion layer. These modifications enhance the backbone network and feature fusion pyramid, respectively. This improves the network's capability to perceive global context information and reduces the semantic gap between fusion features. Experiments on the NWPU VHR-10 and RSOD remote sensing datasets demonstrate that the enhanced YoloX algorithm exhibits a 1.31% mAP and 2.99% mAP performance advancement relative to the initial algorithm. When compared with seven classical algorithms, the recommended algorithm yields a superior detection performance.