DOUNet: Dynamic Optimization and Update Network for Oriented Object Detection

Abstract

Object detection can accurately identify and locate targets in images, serving basic industries such as agricultural monitoring and urban planning. However, targets in remote sensing images have random rotation angles, which hinders the accuracy of remote sensing image object detection algorithms. In addition, due to the long-tailed distribution of detected objects in remote sensing images, the network finds it difficult to adapt to imbalanced datasets. In this article, we designed and proposed the Dynamic Optimization and Update network (DOUNet). By introducing adaptive rotation convolution to replace 2D convolution in the Region Proposal Network (RPN), the features of rotating targets are effectively extracted. To address the issues caused by imbalanced data, we have designed a long-tail data detection module to collect features of tail categories and guide the network to output more balanced detection results. Various experiments have shown that after two stages of feature learning and classifier learning, our designed network can achieve optimal performance and perform better in detecting imbalanced data.