Multi-class Target Detection for Optical Remote Sensing Images Based on Improved RetinaNet

Abstract

In recent years, multi-class target detection in remote sensing images has been widely studied, which is of great importance in both the military and the civil fields. The phenomenon of small targets densely parked (STDP) often exists in such images, people often use oriented bounding box (OBB) method to detect such targets. But the regression of the OBB is difficult, resulting in a decrease in network performance. Therefore, to solve this problem, a cascaded regression module (CRM) is proposed to increase the precision of OBB regression. This paper conducts experiments on DOTA remote sensing data set. Experimental results indicate that the proposed structure can effectively improve the accuracy of multi-class target detection in remote sensing images.