Integration Of Spatial Information Bag Of Feature In Image Annotation

Image annotation is an important problem in computer vision research,a lot of breakthroughs has been made in recent years, popular bag of words model is one of the most important research results, it has proven to be very effective in scene classification and object recognition tasks. Although recognizing objects and scene in an image is still difficult when images include occlusion, scaling change, illumination change and differences between instances, and this task becomes more challenging as the number of categories increase. This paper mainly focuses on these problems and put forward new proposed innovative research, as follows:Firstly, we introduce a user priori based interactive object extraction framework and expand the algorithm to multi-label problems. By combining a multi-graph model and the user brush and the global energy optimization model, we present more fast and accurate object segmentation results.Secondly, when extracting features from images, traditional bag of features model fail to capture image spatial construction. In this paper we add the spatial information into the model, As a result, we not only describe the image appearance statistical information, but also integrate the image local space information into the features .In the end, we demonstrate clear improvements in image scene classification problems over traditional bag of features model.Thirdly, we combine the image segmentation technique and conditional random field optimization, and integrate the spatial feature to solve automatic image annotation problem. We also show how to use image segmentation methods to guard image recognition problem by including the top-down and bottom-up way into a unified framework, and we obtain good results in object recognition task on PASCAL 2007 dataset.We implement a automatic image annotation system and an interactive object labeling system .Given an image, we can automatically locate and identify objects within the image by machine learning and pattern recognition techniques. At last, we present our state of the arts results on those most popular image recognition datasets.