With the easy access of massive high-resolution remote sensing images, it has been an urgent problem that how to quickly and efficiently extract useful information we need from these images. One of the most important problems in this field is vehicle detection from remote sensing images, which has extensive applications, such as military security, urban planning, and urban navigation, etc. Compared with low- and median-resolution remote sensing images, high-resolution remote sensing images provide more details about the earth surface, which enables us to get more accurate vehicle detection results. Therefore, Automatic vehicle detection from high-resolution remote sensing images plays a fundamental role in a wide range of applications. However, although various approaches have been proposed to address this issue in the last decades, a fast and robust approach has not been found yet. As a consequence, it is of urgent demand to develop methods and algorithms for vehicle detection from high-resolution remote sensing images. In this thesis, we analyze the difficulties of object detection in remote sensing images and summarize various methods. To overcome the disadvantages of the existing methods in object search and location, feature extraction and classification in remote sensing images, we propose an efficient method based on deep neural networks. The main contributions of this thesis are highlighted as follows: 1. We propose a simple and efficient approach to automatically detect vehicles in remote sensing images. The proposed model lays emphasis on both speed and accuracy of vehicle detection. It consists of two stages: (1) Thenewly proposed binary normed gradients (BING) is employed to extract reliable region proposals with the purpose of speeding up localization. (2) To enhance the robustness and improve the accuracy rate, we adopt convolutional neural network (CNN) to achieve the goal of feature extraction and classification. Based on the popular open source of CNN framework, Caffe, we design our own CNN architecture. Moreover, the graphics processing unit (GPU) is used to speed up the training of our model. In comparison with the start-of-the-art methods in extensive experiments, we demonstrate that the proposed approach outperforms in terms of both speed and accuracy, thus its effectiveness is verified. 2. We further propose to incorporate sparse constraints into the above mentioned approach for object detection in remote sensing images. Instead o...
修改评论