DETGAN: GAN for Arbitrary-oriented Object Detection in Remote Sensing Images

Siyuan Cheng, Ping Yao, Kai Deng, Li Fu

Abstract: Object detection in remote sensing images has be-come a research focus in recent years with the development of deep learning. However, due to objective reasons such as weather, cost, etc., we can hardly obtain abundant high-quality remote sensing images, especially for specific targets, which severely limits the training of the object detector, leading to poor detection performance. Thus for the first time, this paper introduces the Generative Adversarial Networks(GANs) for arbitrary-oriented object detection in remote sensing images, by augmenting the dataset to improve the performance of detectors. We construct DETGAN with two-layer self-attention modules to capture long-distance dependence for high-quality image generation. To solve the mismatch between generated slices and the samples for detectors, we propose the GAN-to-Detection transfer strategy, in which the slices are inserted into a background with the same size as the samples for detectors and then added to the training set. Experiments show that the performance of ship detectors is successfully improved with the transfer strategy, and demonstrate that GAN is an effective way to alleviate the problem of data insufficiency in remote sensing image object detection.