Image Descriptors And Feature Retrieval In High Dimensions

The task of organizing all of the visual information available to us today is overwhelming. Digital imagery has become increasingly widespread thanks to the popularity of mobile camera-phones and online photo-sharing services such as Flickr1 and Picasa2. Content based image retrieval (CBIR) can return more accurate and visual images to users relative to key-words based image retrieval. So CBIR has been one of the hottest topics in pattern recognition, computer vision and artificial intelligence. We mainly focus on sparse features and 3 excellent frameworks for feature retrieval in high dimensions.As to the image descriptors, the global features are reviewed for short. The paper mainly focuses on the sparse feature based on interest point detection. How are the interest points detected with scale invariant detectors and how can the detectors solve the scale invariant problem are introduced firstly. And then, SURF and SIFT descriptors are taken as examples to introduce how the descriptors are constructed to solve the problem of rotation and illumination changes.In the part of feature retrieval in high dimensions, Local Sensitive Hash (LSH), Multiple Instance Learning (MIL) and Vocabulary Tree (VT) and some improvement works based on the 3 frameworks are reviewed. The paper also analyzes how can they retrieval the approximate KNN fast and accurately based on their rationale.The paper focuses on our innovative works. We propose to construct adaptive vocabulary tree based on the average variance of the node, which can provide more suitable hierarchical classification. And we also propose a novel tree fusion framework: Feature Forest, utilizing and fusing different kind of local visual descriptors to achieve a better retrieval performance. We realize the adaptive vocabulary tree and the feature forest based on HOG vocabulary tree and HOG vocabulary tree. The adaptive vocabulary tree is tested with UKY database and the result of the experiment shows that the precision of its retrieval is better than standard vocabulary tree. The feature forest is tested on UKY, ZuBuD and famous_landmark databases, on which the retrieval difficulties are different. The retrieval precision of the feature forest is always higher than the HOG tree and SURF tree in each of the databases. So the evaluations show the effectiveness of our approach compared with single vocabulary-tree based on different databases. The system is realized based on c++ and OpenCV.