Research On Local Invariant Features Of Image

It is difficult for computer to perceive and recognize the objects in nature. It’s essential to select appropriate feature to represent the scene. Not only should feature distinguish different objects, but also overcome the problems which are brought by various transformations. Feature detection is an essential component of computer vision. According to the local information of image, researchers have made use of multi-scale analysis and statistics approach to represent image content. And these methods have been widely used in object recognition, image registration and image stitch.Each local invariant feature has different characteristic and its appropriate usage. It is important to choose the proper feature detector for specific application. Firstly, the background of this topic is introduced in this paper. After summarized the related work about local invariant feature, we have described the algorithm, mathematic model as well as implementation of corner, blob, region detectors in three chapters. The common applications for each method were also respectively listed.Considered the application for feature detectors, several important invariant properties were proposed. We focused on testing the repeatability and matching scores of each feature detector. Furthermore, to evaluate the performance of detectors for object recognition, we have tested the average precision on clustering. Based on pyramid match kernel, we have used SVM classifier to compare and analyze recognition result for different feature detectors. What’s more, it was concluded that several principles for choosing appropriate local invariant feature. However, there is still a lot of space for improvement in the future, as several problems haven’t been solved yet.