Multi-scale Feature Enhanced Network for Remote Sensing Object Detection

Abstract

The task of remote sensing image target detection still faces many challenges: the target scale, orientation and spatial distribution are very different. In particular, small targets are difficult to identify in complex backgrounds, and the change of target geometry will affect feature extraction and positioning accuracy. In view of the above problems, this paper builds an improved detection network MSTD-Net based on YOLO26 m. Firstly, the global context Transformer module (GCTM) is introduced into the backbone network to capture the context information between long-distance spatial regions. Secondly, the dynamic deformable convolution module (DDCN) is used to adaptively adjust the convolution sampling position according to the geometric shape of the target, so that the network can better adapt to the targets with different shapes and orientations. In addition, this paper uses the dynamic Wise-IoU loss function (DWL) to adjust the contribution of training samples with different positioning quality in the bounding box regression process. In this paper, a large number of comparative experiments are carried out on DOTA15, UCAS-AOD and HRSC2016 datasets. The experimental results show that the mAP @ 50 index of MSTD-Net on the three datasets reaches 7575 %, 9730 % and 7403 %, respectively. Compared with the baseline model YOLO26m, MSTD-Net increased mAP @ 50 by 100 %, 026 % and 792 % on DOTA15, UCAS-AOD and HRSC2016 datasets, respectively. The ablation experiment further verifies that the three improved modules all have gains in detection performance, but the improvement is different on different data sets.