Reppoints-Based Multiscale Task Enhancement Network and Sample Assignment Method for Oriented Object Detection

Unlike normal images, remote sensing images (RSIs) often contain complex backgrounds and multiscale targets with arbitrary directions. This makes the existing detection methods ineffective. In contrast to the usual rotating frame RSI target detection methods, the Reppoints-based method can learn autonomously to capture target features of arbitrary pose based on the target's characteristics. Therefore, a new Reppoints-based detector is proposed in this letter to improve the accuracy from multiple perspectives. To better expand the receptive fields and obtain finer features, the multiscale task enhancement network (MSTEN) preserves the multiscale receptive fields and multiscale information through the deformable convolution (DCN) and the skip connection, while improving the feature extraction capability of the network for targets of arbitrary orientation. At the same time, the network adapts the features to the characteristics of different tasks to better suit the needs of the respective tasks. Finally, the dynamic reppoints learning (DRL) is proposed to select samples that perform well in both the regression and classification tasks. The experimental results on two challenging datasets, DOTA and HRSC2016, show the effectiveness of our proposed method.