Recognition Of Multi-class Objects Based On Intra-category Shared Features Model In Remote Sensing Images

With the improvement of the satellite optical imaging resolution, more details and categories of targets that can be recognized are provided by remote sensing images. Thus, algorithms of higher efficiency and more robustness are needed in detection and recognition of multi-class targets. With the purpose of improving the recognition rate of multi-class objects in the remote sensing images, our work in this paper, includes feature extraction of multi-class objects, feature modeling of multi-class objects based on shared-feature, and recognition methods of multi-class objects consisting of airports, harbor, bridge, highway, railway, island and oil deposit.Firstly, we extract some simple features such as parallel line and L-junction, and obtain the image primitives by features combination. Then an efficient description of object can be obtained by joining the image primitives and the structural information. Meanwhile, we use AdaBoost for feature selection. In this way, the obtained object feature model is more stable as a result of analyzing the contribution of different features to the classification task.Secondly, the feature model of multi-class objects is built by analyzing features of objects of interest, fusing the idea of multi-resolution and feature quantification and combining salient features of each object and shared features of multi-class objects. Since they have different scale sizes and efficiently stable features, we divide the interest objects into several subclasses which have shared features.Finally, the paper proposed a multi-class object recognition algorithm based on shared features. Features with high saliency, stability and capacity in classification will be chosen prior. Meanwhile multi level and multi class object detection and recognition is performed by combining shared features. The efficiency of computation will be improved by the scene contextual information of objects and the interest regions obtained according to the density of lines. Experiments demonstrate good performance in detection and recognition of multi-class objects in remote sensing images.