Panoramic image construction using feature based registration methods

This paper presents the method to stitch multiple photographic images having overlapped fields of view to construct a panorama. Panoramic image provides wide view and covers the various aspect of scene under consideration. Keypoints from captured images have been extracted using Scale Invariant Feature Transform (SIFT) and Speeded-Up Robust Features (SURF). These input images vary in factors like illumination, viewpoint, rotation, scale, blurriness and noise (i.e. Gaussian and salt and pepper). Image registration geometrically aligns the images taken at different times which vary in their illumination. The heterogeneity in image illumination effects the feature extraction and ruins entire stitching process by leaving a visible seam at the junction of degraded images. This anomaly is overcome in preprocessing by making the intensity levels consistent in HSI color domain and also with the histogram equalization. Another widely practiced way for seam removal is to blend both images at the time of image stitching. Keypoints given by feature extraction method contain mismatches or outliers which are further removed from data using Random Sample Consensus (RANSAC). Homographic matrix has been estimated and used to implement affine and projective transformation model. Detailed discussion on the results is given at the end by comparing the techniques side by side. Potential applications for panoramas are medical imaging, high resolution photography and painting reproduction, real estate virtual tours, cinema etc.