FPN-GAN: Multi-class Small Object Detection in Remote Sensing Images

Abstract

Despite the recent dramatic advances in object detection, detecting a small object in general and in remote sensing images is still a challenging problem. One main reason for this is the appearance of small objects in images. Specifically low resolution and noisy representation makes it hard to detect small objects. We tickle down this problem by proposing a novel object detector based on Generative adversarial network (GAN), which we called FPN-GAN in short. The proposed method is composed of GAN, Resnet-50 as a backbone, and Feature Pyramid Network for detection. We combine both of these methods to achieve a single end to end GAN model for multi class-small object detection and image enhancement simultaneously. Extensive experiments on a challenging benchmark DIOR remote sensing dataset demonstrate the superiority of the proposed method for small objects as well as large and the medium size objects.