Saliency Detection Of SAR Images With Fully Convolutional Neural Networks

Salient object detection of SAR images is a challenging task but benefits many downstream tasks of SAR image interpretation.The previous works tackle this problem by bottom-up contrast analysis with hand-crafted features.Early bottom-up models are designed based on hand-designing features such as intensity,edge orientation,color,or center-surround contrast.These models are designed for specially for the optical images.Optical images captured in the natural scene have few conspicuous objects at the center of the images and these objects are easily distinguishable from the background.Hence,these bottom-up traditional methods work well on optical images.Whereas,SAR images have different imaging principles and characteristics.Unlike natural images,SAR images are taken from longer distances and these contain remote sensing information of multiple objects scattered in the scene,moreover,these images do not have color and center-surround contrast.Hence traditional methods are not a good choice for the saliency detection of SAR images.In this work,we present a novel end-to-end SAR Saliency Network(SSNet)based on fully convolutional neural network to predict the saliency map of SAR images.Our proposed method is a top-down method that learns the saliency map which is a continuous distribution,in which the brightness of each pixel indicates the probability of human eye fixation on that pixel generated by human gaze.Experimental results on the collected real world SAR-Gaze dataset demonstrate that the proposed deep-learning based model,SSNet,accurately predicts the saliency map of SAR images with least computational complexity.Moreover,the proposed network also mitigates the annotations noise of eye tracking device in its prediction.