Research On Object Detection And Recognition For Visible Remote-Sensing Images By Introducing Context

Object detection and recognition in visible remote sensing images is a very important aspect of remote sensing technology. The characteristics of objects often vary largely due to the influence of different imaging conditions. The difficulty of object detection and recognition is increased greatly. The efficiency and performance could be improved by using contextual information. This dissertation focuses on introducing contextual information for object detection and recognition in visible remote sensing images as follows.We investigate object detection in the scene in which object’s close background is relative stable. To address the disturbance of the false alarm that is very similar to true object, we propose a detection framework by introducing neighbor context. In the framework, we propose a contextual feature descriptor called Binary Pattern of Oriented Gradients. Due to the needlessness of code book learning, the efficiency is enhanced significantly. Also, we apply this descriptor in spatial pyramid matching model, and then the object detection performance is improved.We investigate object detection in variable scene and propose an object detection framework by introducing object context. For objects’ contextual relationship, there sometimes exists the puzzledom that the same relationship may occur between different categories and different relationship may occur in the same category. To solve this problem, the objects’ co-occurrence relationship is converted into contextual words and used to train a probabilistic Latent Semantic Analysis model.We investigate object category recognition and propose a framework by introducing internal context of object. There often exist errors during feature extraction. To avoid this problem, we start with extracting sparse salient features from object. Then, we combine internal position information with features to classify objects. The robustness of feature description and the performance of recognition are improved.We investigate land-cover classification of remote sensing images. Since the object’s characteristics of spectral, shape and texture features are complex in the same category, we propose a land-cover classification framework by introducing multiple contexts. In order to extract good segmentation object, a new hierarchical object segmentation approach based on graph model is proposed for object acquisition. Then, a rotation-invariant representation of neighbor context is proposed to refine the initial classification for segmentation object. Finally, Markov Random Field(MRF) model is used to introducing contextual constraint between objects. The proposed framework employs a hierarchical strategy to introduce multiple neighbor contexts and solve the difficulty that the MRF model relies on initial classification results.