Multilevel color histogram representation of multiresolution images by peaks: Room Recognition System

Designing an accurate, robust, and fast indexing method to work on a wide range of imagery is a formidable challenge. Color-based image indexing is an important approach because of it's easy to compute, low-level and task-independent definition. In our work, we are interested in developing an very fast color-based indexing scheme. The work in this dissertation contains the following contributions: (1) A novel approach to color image indexing using histogram peaks. (2) Extensions to multilevel histogram peak indexing. (3) Evaluation of a system using this techniques for object recognition.; Peak matching provides a new approach for the application or multilevel color histograms for object recognition. Compared to past techniques it is more stable under scaling changes and lighting changes. A multiresolution algorithm by peak indexing for finding scale effects is presented. This algorithm is able to achieve good results on images at different scale. To our knowledge, we are the first to combine a multiresolution/image pyramid approach will) color histograming. The problem of color histogram sensitivity from an image resolution has been investigated. We also studied the differences between smoothing color histogram of images and color histogram of smoothing images characteristics for scaling process. The success of the overall system is due to several novelties, including an intensity variation constraint and a histogram peak similarity measure that is relatively insensitive to image resolution change.; A summary of peak-based recognition comparison measurement has been given, along with a verification of their success for objects in the COIL-100 database which includes 7200 images. It was also compared to a range of other histogram-based techniques. The proposed peak-based techniques yielded a recognition rate of 100% on this data set when looking from an object from its training set. Measurements on robustness to viewpoint variations are also discussed.; Finally, a real-time algorithm to recognize a room in an unmodified complex environment has been presented as an example “Room Recognition System”. Unlike conventional techniques, the algorithm is insensitive to full 360-degree view of camera, daily light change, arbitrary camera movement, and multiple moving people in the background, all simultaneously. It provides >92% recognition rates across a database of hundreds of rooms.; In summary, the main motivation of histogram peak matching for object or room recognition in real-time is its low computational cost and less storage requirements. While a very compact representation the experiments in this thesis suggest it is a viable approach.