Boosting Small Ship Detection in Optical Remote Sensing Images via Image Super-Resolution

Small ships in optical remote sensing images are hard to detect due to the lack of sufficient detail information. In this paper, we adopt the image super-resolution technology to solve this problem. Specifically, an effective super-resolution network is designed to generate clear super-resolution ship images from small blurry ones produced by the ship detector. Inspired by the idea of generative adversarial network (GAN), the super-resolution network is trained together with a discriminator network in an adversarial way, aiming at generating more realistic super-resolution images. Moreover, to eliminate false detections, the discriminator network is also used to distinguish ship and non-ship images via an additional classification branch. Experimental results demonstrate the effectiveness of the proposed method.