MA-Net: Multi-Scale Adaptive Network for Oriented Object Detection in Remote Sensing Images

The remote sensing image detection task faces significant challenges, e.g. complex background, dense distribution, widely varied target scale, and diverse target directions. This paper proposes a remote sensing image Oriented object detection network (MA-Net) that integrates parallel multi-scale feature extraction and adaptive dynamic label assignment. The network is based on the FCOS object detection algorithm, which introduces a parallel multi-scale feature extraction module to enhance the model's understanding of multi-scale objects. The adaptive dynamic label assignment algorithm is introduced to enable the model to select samples that are more worthy of learning. According to the number of positive samples at different levels, hierarchical supervision reweighting is used to help the model focus on more important levels and improve detection accuracy. Experiments are conducted on two public datasets, DOTA and HRSC2016. Compared with the original FCOS algorithm, the detection accuracy of our algorithm was improved by 2.45% and 4.44%. And its performance is comparable to that of state-of-the-art detection methods.