Local Feature-based Object Detection Research

High resolution optical remote-sensing images have been widely used in many civilian and military aspects. This paper mainly focuses on the methods of detection of small targets such as military ships and aircrafts. From general object recognition methods to recognition methods of targets in remote-sensing images are discussed. And from their common points and the particularity in remote-sensing images, two algorithms based on local features were proposed to deal with the problem of ship and plane detection respectively.For ship targets, an algorithm based on image segmentation and detection of the bow of the ship was proposed. In high resolution remote-sensing images, the shape of ships are similar to each other, and the difficulty lies in the detection of adjacent docked ships and ships docked along the shore. We supposed the water field has been extracted from the port, and ignored the problem of separating the port and the ships. After a coarse segmentation of ships was done, a method to detect corner point and angle was used to determine the bows of ships and to separate the ships docked adjacently.For plane targets, a few of shape templates were used for different types of planes. In order to remove the effects of difference in contrast, color and a small amount of texture, local self-similarity descriptors were used to describe local shape features. A speed-improved belief propagation algorithm was used to match the global features. Experiments shows this method is effective and can not only be applied to the detection of intact planes, but also can be applied to the detection of shape-incomplete planes and finding the damage part of planes.