Superpixel Segmentation:Theory,Algorithms And Applications

Superpixel segmentation,feature extraction and image stitching are very exciting topics in computer vision.This thesis work revolves around these areas.The digital image is a kind of data structure,having certain number or codes for each and every pixel or alternatively called picture element in the image.This number or code for each pixel determines its color.Each image pixel is a discrete sample of a continuous real image.However,contrary to the concept of image pixel grids,superpixels are perceptually uniform and computationally efficient regions or clusters of pixels as.Superpixels are the result of over-segmentation process.Since superpixels reduce the computational cost and redundancy to a great extent.Our work focuses on segmenting a pair of images into superpixels first and then using superpixelized images for feature matching,registration,alignment and stitching to produce panoramic image.Superpixels have been used in many computer vision applications and it is observed that they can bring interesting results when used for image stitching.Therefore,image stitching is also addressed in the thesis using superpixels rather than pixels.A superpixel based image stitcher is designed to provide a substitute to acquiring panorama images with a panoramic camera.In essence,two very familiar approaches to image registration are direct(pixel-to-pixel based)and feature-based approach,we have used later approach for image stitching in our work.Looping over the superpixels reduces time and complexity,as there are very few superpixels of an image when it is oversegmented.Moreover,direct pixel-to-pixel methods are not feasible for feature extraction in superpixelized images.SLIC algorithm is used for extracting superpixels,for feature matching we have used the SIFT method and for inlier selection,RANSAC is used.Finally,images are placed onto a composite surface using some warping transformations.It is also shown that our image stitcher works very well for both natural and synthetic images.SIFT based matching using superpixels results in quite less but interesting number of features that are adequate for matching two images.Runtime of superpixel based image stitching is less than the time consumed to stitch original images.Images used in our experiments are taken from PASSTA(Panorama Sparsely Structured Areas)dataset,which contains challenging natural and synthetic images.