Saliency-based End-to-end Target Detection Model in Optical Remote Sensing Images

Abstract

It is challenging work that detecting target such as aircraft in remote-sensing images because of the complicated background. Most existing methods are based on deep-learning network considering its high learning power of the features. However, the region proposal network is often based on deep network, which cost much time and computation on the proposal of the irrelevant regions which are useless to the target detection. Based on the above considerations, a novel end-to-end aircraft detection model based on saliency map is proposed in this paper. The saliency-based region proposal network can produce the target-like regions and filter out the most irrelevant background regions. Meanwhile, it cost less computing time compare to the network based on deep-learning network. Then, a novel target detection network is designed to extract the feature of target-like regions, and classify these features by the iterations of a coupled networks, the result of the binary classification is conducted by the classification layer, at same time the accurate bounding boxes are conducted by the regression layer. The performance of our method is evaluated by detecting aircraft targets in high resolution remote-sensing images. Superior experimental result proves the effectiveness and efficiency of proposed model.