A vector sift operator for interest point detection in vector imagery and its application to multispectral and hyperspectral imagery

This research work presents an algorithm for automated detection of interest points in vector images such as RGB and hyperspectral. Interest points are features of the image that capture information from its neighbors are distinctive and stable under transformations such as translation and rotation. Interest point operators for grayscale images were proposed more than a decade ago and have since been studied extensively. These operators seek out points in an image structurally distinct, invariant to imaging conditions, stable under geometric transformation, and interpretable. Interest points are helpful in data reduction, and reduce the computational burden of various image processing algorithms. The developed approach, extends ideas from Lowe's operator that uses local extrema of Difference of Gaussian function at multiple scales. A modification to Lowe's approach to vector images is proposed. The multiscale representation of the image is generated by vector anisotropic diffusion that leads to improve detection since it better preserves edges in the image. Vector ordering methods are used to find local extrema and second fundamental form is used for curvature analysis to eliminate poorly defined extrema. Experiments with RGB and hyperspectral images to study invariance to translation, rotation and scale changes are presented. The performance of the detector is quantified using repeatability criterion and image registration.