A hybrid framework for efficient and accurate orientation estimation based on single and multiple orientation vector fields

Abstract

This article presents a hybrid framework for efficient and accurate orientation estimation. The proposed scheme combines the single orientation information given by a novel method and the multiple orientation information provided by a bank of linear orientated morphological openings. The single orientations are estimated by means of an energy-minimization Gaussian filtering which solves the drawback related to phase changes of other methods. After describing the formulation of these two approaches for estimating the existing orientations in the pixels of an image, several strategies have been analyzed to fuse and discriminate the information of both orientation vector fields in the resulting hybrid orientation vector field. The objective of the proposed hybrid method is to reduce the computational cost involved in calculating multiple orientations only in those pixels where they exist while maintaining the accuracy provided by the single orientation method in the remaining pixels. To this end, strategies ranging from a threshold in the multiple orientation vector field to a convolutional neural network trained with a set of patterns specifically designed to detect pixels with multiple orientations, passing through the Harris corner detector, have been tested to identify those pixels where multiple orientations exist. Results on natural and synthetic images show the accuracy and the computational efficiency achieved by the proposed hybrid framework to provide the vector field with single and multiple orientations.