Multi-Class Object Recognition In High Resolution Remote Sensing Images

Regarding the multi-class detection and recognition task in high-resolution remote sensing images, the method used in this dissertation is different from the traditional methods that build specific model for each class. The objectives of the dissertation lie in two folds: to detect and recognize multi-class objects in large scale images by extracting Regions of Interest (ROIs), and to obtain high recognition rate as well as reduced recognition time.Aiming to detect and recognize five classes of man-made construction: airport, harbor, bridge, highway and railway, geometrical features were used to extract ROIs and feature vectors, texture classification results were referred to judge context information, tree recognition structure was adopted in multi-resolution. The key points are: to generate validate feature vectors to depict each class, to extract ROIs to eliminate the background interferences, and to design effective classifiers to recognize potential objects.The main contributions given in this dissertation are as follows:Firstly, as the man-made constructions show lots of line segments, line segments were extracted as the basic features. The gradient direction group line segment extraction algorithm was improved by referring to the edge information obtained by Canny operator. Proved by the experiment results, this improved algorithm is with high efficiency, robust to noise, and able to extract the key line segments.Meanwhile, the structural matching and Bag of Features ideas were integrated to generate feature vectors. Based on the basic line segments'shape, geometrical and spatial relationships, the shape primitives and their attributes were calculated and the feature vectors were constructed.Secondly, as the man-made constructions represent dense line segments, a novel ROIs extraction method was proposed for man-made constructions in remote sensing images, the ROIs were obtained by calculating line density distribution surf. This method is proved to be able to extract the man-made regions excellently.Finally, the parameter optimization in designing SVM classifiers was discussed, and a method of multi-class objects detection and recognition in high resolution remote sensing images was proposed. The experiment results showed that a relatively high recognition rate and efficiency was obtained.