Refined regression detector for multiclass-oriented target in optical remote sensing images

Abstract

In recent years, multiclass target detection in remote sensing images has become a popular research topic, and it has been widely applied in military and civilian fields. Multiclass-oriented target detection in remote sensing images presents the following challenges: small densely parked targets (SDPT), multidirectionality, interclass unbalanced number of samples (ICUNS), and hard example detection. These problems will impact the result. Therefore, to solve the abovementioned problems, we propose a multiclass-oriented target detection method in optical remote sensing images. In the detection stage, an oriented bounding box (OBB) is used to predict the angle of the target, which can overcome the problems of SDPT and multidirectionality. A cascade refined module is proposed to solve the problem of network performance degradation caused using the OBB. Second, the smooth L1 loss function is used, which can complete OBB regression by adding an angle parameter. This method can improve network performance. Finally, gradient harmonized mechanism loss is applied to the OBB. It can solve problems, such as ICUNS and hard example detection. We describe experiments conducted on the DOTA public optical remote sensing dataset. The experimental results show that this method is effective for multiclass-oriented target detection in optical remote sensing images.