Research On Flip Invariant Feature Description And Matching Method

Image understanding is based on image feature extraction, description and matching. All kinds of image features can be grouped into two categories, i.e. global image feature and local one. Global image feature takes entire images as units to extract properties, such as color statistic histogram and Euler number. As for local image feature, its speed is inferior to the global one, however, it is very robust to most of geometric and photometric transformations. And time expenditure is not a bottleneck problem any longer, due to the benefit from hardware development. Therefore, the local image feature is the most popular technique for image understanding currently. This thesis presents a novel concept, Mirror-reflection Invariant Feature Transform (MIFT), for SIFT-like descriptors to improve their degenerated performance due to mirror reflections. Based on MIFT, the author proposes a neat Flip INvariant Descriptor (FIND) with consideration of spatial relationship.Besides feature extraction and description, feature matching is another critical part in the framework of image understanding. A good matching scheme has two main factors, i.e. matching accuracy and number of correct matches. To simultaneously improve both matching accuracy and correct match amount, this work proposes a measurement based on Triangle-Constraint (T-CM) for providing local image feature based applications with more valuable information.