A Novel Oriented Object Detection Method Based on Multi-Scale Feature Fusion and Diagonal-SmoothL1 Loss in Remote Sensing Images

Currently, object detection has been a fundamental task in the computer vision. In remote sensing images, objects with different scales and orientations hinder the existing methods from achieving higher performance of object detection. To achieve high-precision oriented object detection, we propose a novel oriented object detector based on multi-scale feature fusion and diagonal loss in remote sensing images. The multi-scale feature fusion module is designed to effectively integrate different scale features from various levels of backbone. This enables the final features at each scale of feature pyramid networks (FPN) to contain both low-level object information and high-level semantic information, which helps the network achieve more accurate object classification and localization. Besides, some existing methods usually require additional processes to generate rotational proposals from horizontal anchors for avoiding the direct regression of angle parameters in the region proposal network (RPN), which can potentially induce the position errors and decrease the performance of object detection. To mitigate the influence of additional processes, we construct a new Diagonal-smoothL1 loss (DS-Loss) by combining the diagonal loss and the general smooth-L1 loss, further improving the accuracy of object detection in remote sensing images. The experimental results on two public datasets (HRSC2016 and DOTA) demonstrate that our method can outperform other state-of-the-art methods.