Small Object Detection in Complex Large Scale Spatial Image by Concatenating SRGAN and Multi-Task WGAN

With rapid development of IOT, especially in the field of geological exploration, areal images obtained through optical sensors with large spatial coverage are becoming an effective material for earth understanding. As such that research on object detection among large scale spatial image is of great significance for resource exploration, natural disaster assessment, and military target detection and recognition. Compared with common images, object detection in spatial image is still a big challenge because of the spatial heterogeneity, scale dependency and with complex context. In order to solve this problem, we propose a novel framework by concatenating a super resolution GAN (SRGAN) and a multitask Wasserstein GAN (WGAN) to increase the performance of a backbone detection model, like RetinaNet. Here, SRGAN and WGAN act as image restoration module attempt to increase global and local resolution respectively. To be exactly, the SRGAN is applied to the global image before detection, while the WGAN is applied to the anchor regions after detection. Unlike previous method with unique generative task, our WGAN is designed to complete generation, classification and regression tasks, with which can increase the detection performance comprehensively by back propagating classification loss and regression loss to the generator to remove false positive anchors and regulate the region box. In order to verify the effectiveness of the proposed model, we conducted a large number of ablation and comparative experiments on the public areal image dataset DOTA, and obtained superior average precision for small object detection.