The Study And Application Of SIFT And BIM Feature

In the Information Age, our society is in great demand of effective methods of categorizing tons of images data. Object recognition/categorization, whose task is to search for effective and efficient ways in image categorizing, become one of the most popular and important research topics in computer science.In the system of image categorizing, feature extraction module is considered as the most vital part. Therefore a lot of study and research have been put into finding how to extract good image feature. Many successful achievements are obtained and lots of delicate feature extraction algorithm designed. The most representative algorithm among them, are the SIFT and BIM algorithm. The thesis study on these two algorithms, including the theory and implementing details of SIFT and BIM, the application and improvement of SIFT in offline handwritten character recognition, the application of BIM in facial expression recognition.First, we studied the Scale-Invariant Feature Transform, known as SIFT, including the theory of its design and the implementing details. SIFT algorithm containing two parts:the SIFT detection and the SIFT descriptor. The SIFT detection algorithm is an algorithm base on scale space theory, it use Difference of Gaussian space to locate feature; and SIFT descriptor is a data structure describing image region base on gradient histogram.Then, the application of SIFT in handwritten character recognition was studied. We proposed two novel SIFT-based features for offline Handwritten Chinese Character Recognition (offline HCCR). The first feature, named Local Character-SIFT, uses elastic meshing to segment character image into patches, then extracts series of SIFT descriptor from these patches. The second feature is called Global Character-SIFT, which first constructs global elastic mesh to form several sub-regions and then accumulates the related gradient code of each sub-region. Experiments show the merits of our new feature while comparing to original SIFT feature, Gabor feature and gradient feature. Moreover, we point out that the character-SIFT feature has scale parameter and we study performance of the merging among different scale character-SIFT feature.Finally, we focused on Biological-Inspired Model, known as the BIM feature. BIM feature is a hierarchical system that closely follows the organization of visual cortex, the theory and the implementing details of it were closing studied. Then, we applied BIM feature on Facial Expression Recognition (FER), and studied the effect caused different parameter of BIM. Our experiment shows that BIM feature has good performance in FER.