Deformable Templates Based On The Combination Of Alignable And Non-alignable Skeches

The recognition of object is a very active branch in the image processing and computer vision field. Research in this domain involves wide applications, including security surveillance, human-computer interface and the details analysis of human movement. Deformable templates and sketches are both important elements in object recognition. Sketches are chosen for object recognition because that is a very simple and intuitive way to represent object information. Recently, Wu et al proposed Active Basis model for deformable templates to share sketches, where each sketch is allowed to shift in position and orientation.Based on the summary and analysis of the relevant research, this paper proposes a hybrid model for deformable templates which combine alignable and non-alignable sketches. These sketches are subject to slight or considerable translations in different images. For slight translations, Wu et al proposed active basis model to capture them, where each sketch is allowed to shift in position and orientation. For larger translations of sketches, Wu et al assumed that they follow the same distribution as sketches of natural image ensembles, which need not be explicitly modeled. But in fact, for a specified object class, the unaligned sketches follow a totally different distribution from those of natural images. This thesis summarizes these sketches by their means in the foreground mask and treats the mean value in each direction as independent features. Moreover this thesis fits their marginal distributions on object ensemble and natural image ensemble using Gaussian distribution. The marginal distributions are combined with Active Basis into a joint probability ratio to distinguish foreground object from natural background. Compared with original Active Basis and its variants, experiments show that this new model significantly improved the recognition performance and can recognize objects with cluttered background.Experiments are conducted on 50 object classes. First of all, this thesis trains corresponding Active Basis model and new model for every object. Then experiments are separately carried out in testing samples using the two models and draw their ROC curves. The results show that the recognition capability of the new model considerably improved the performance in ROC.