Superpixel Level Object Recognition Under Local Learning Framework

With the wide applications of computer vision and the development of image processing techniques, increasingly high standard of automatic image content identification by computers is required, to meet the need of the real world applications and higher level image processing schemes. Image object recognition is not only required to make sure whether an image contain an object category, but also asked to get the location and the scale of the object.Traditional sliding window object recognition method finds the object region by scanning the whole image with a bounding box, which results in large computational cost while only obtains a coarse possible location of an object. Recognition methods based on segmentations are segmenting the image into regions first and then identifying each segment with object classes. However, these methods usually rely too much on the accuracy of the segmentation results. Pixel level object recognition aims to assign a category label to each pixel of an image, which suffers from high computational cost and data redundancy as well as low special consistency.This paper utilizes superpixel as the elementary unit of the categorization and localization methods, and launches the object recognition task by evaluate the probability of each superpixel belonging to some image category, since superpixel is good at capturing the local redundancy and preserving the boundaries of an image. To solve the intraclass variations problem of the superpixel classification, this paper proposes a local learning framework to turn a highly non-linear classification problem into multiple local linear problems within different subset of the database. Besides, we integrate each superpixel with its neighbors within neighborhood distance N to pursue spatial consistency, and correct the integration with superpixel mean color map to make a clearer boundary around the objects. At last, we utilize superpixel based Graph Cuts algorithm to segment the objects from background image.We evaluate our superpixel level object recognition method on the challenging Graz-02 dataset. Experiment results demonstrate that our method outperforms the state-of-the-art methods and can get object segmentation results with clear boundaries.