PROJECT TITLE :

Online Subspace Learning from Gradient Orientations for Robust Image Alignment

ABSTRACT:

Robust and effective picture alignment remains a difficult task due to the size and complexity of images as well as fluctuations in illumination, partial occlusion, and corruption. We present an online picture alignment method based on subspace learning from image gradient orientations in order to address these problems (IGOs). An online framework that is robust for aligning images with significant intensity distortions incorporates the proposed method's subspace learning, modified IGO reconstruction, and image alignment into a single online framework. PCA based on gradient orientations provides a more trustworthy low-dimensional subspace than does PCA based on pixel intensities, which is the inspiration for our approach. The IGO-PCA basis learned from previously well-aligned images can be decomposed as a sum of a sparse error and a linear composition of the IGO-PCA basis instead of processing in the intensity-domain as in standard approaches. In order to solve the optimization problem, an iterative linearization is used to minimise the _1 -norm of the sparse error. A thin singular value decomposition based on the change in IGO mean is used to adaptively update the IGO-PCA basis. Image alignment, medical atlas generation, and face recognition are used to demonstrate the effectiveness of the proposed method on a wide range of hard datasets. According to our experimental results, we have developed an algorithm that is more robust to illumination and occlusion than other techniques.