Rotation-Invariant Descriptor for Disparate Images Using Line Segments

Abstract

In recent years, computer vision applications have extended to a very wide range, which in turn encompasses a large variety of situational images and videos. This paper modifies the Duality Descriptor (DUDE), which uses line-point duality that provides simple consistent method of feature extraction. DUDE descriptor works very well for disparate image pairs, often outperforming most other methods with significantly less computation expenses. However, DUDE descriptor is not invariant to scale and rotation changes to the image, which is often vital for image processing in real-time scenarios. This paper modifies the existing DUDE descriptor, making it invariant to rotation to a certain degree. The experiment has been performed for some real-time images of objects to show the viability of the proposed descriptor. Herein, multilayered neural network is also used to verify the results in terms of percentage accuracy.