Aircraft Segmentation from Remote Sensing Images using Modified Deeply Supervised Salient Object Detection with Short Connections

Image segmentation is a process of clustering the pixels into some groups. It is essential and useful in image processing and computer vision task. The segmentation result may affect the outcome of many types of work, for example, image recognition and image detection. One of the most exciting works of image segmentation is aircraft segmentation from remote sensing images. This work is very challenging due to many factors, such as the complexity of the background, the shadow, and the various type of aircraft. To handle these problems, in this paper we implement the deeply supervised salient object detection with short connections (DSS) with our dataset which is divided into two categories: simple case and dificult case, and we propose some idea of post-process which can increase some accuracy by using pre-trained and adding L2- normalization in some layer of DSS architecture to create the mask that can reduce the noise from background effectively. From the experimental result, using L2- normalization with our post-processing can increase the performance by 0.61% and 2.14% in the simple case and difficult case respectively, compared with using only the DSS method. So our method performs well when the background is complicated, which always occurs in reality.