Research On Object Localization And Detection In Digital Images

With the rapid increment of images on World Wide Web, how to query and classify images efficiently becomes a hot research topic. Object detection methods studied in this paper is just this kind of work:extracting object information from digit images. Basically, Object detection includes two parts:object classification and localization. What we introduced here covers the following aspects:low level feature extraction, feature selection and object detection. We have proposed some new methods on these aspects based on existing methods, and tested them with experiments.This thesis presents our works on following aspects:Firstly we introduce different feature extraction methods, feature representation, feature comparison and selection, and feature matching methods. On the part of feature matching, we introduce our work and contributions.In the research of ROI based object detection method, we improve the state-of-the-art feature selection framework by using foreground/background contrast information. Then we use these words to do image classification. Compared with current methods, our selected feature is more representative and powerful. Object detection precision can be improved by using new features.In the research of FSESS method, we accelerate ROI detection method by efficient sub-window search (ESS). ESS is a divide and conquer framework of sub-windows searching Based on the ROI detection method, we first theoretically prove that our work can be solved by ESS framework. Then we prove the efficiency of our work by experiments on real images.In the research of segment map method of object localization, we discuss the relationship among pixels, segments and object areas. Segment map method is a bottom up method that combining low level pixels and segments. Experiments on PASCAL VOC2007dataset show that this method is an effective and efficient object localization method.In the research of saliency and defocus based foreground extraction method, we find that blurriness of foreground and background are different in image especially those taken by professional photographers. This information can be used to separate foreground from background. Based on this idea, we propose a new definition of foreground:areas near and clear. We propose a method combining saliency based method with defocus method to extracted foreground areas. Experiments show that our definition is reasonable and our method is effective.