Method of smart corner detection developed with artificial neural networks

Feature detection is an important part of computer vision. It is difficult to detect invariant features in different scales and transformations, in varying brightness, small angle changes, noisy, mixed and worn images. Scale Invariant Feature Transformation method aims to detect unchanging features and in doing so it goes through four stages. It uses the corner detection method in the stage where the key points are located. The corners are very useful features in image processing techniques which are often used in object recognition, motion tracking and stereo matching methods. It is possible to use the corners to identify, distinguish and determine the object. Methods for performing corner detection in a gray-level image can be grouped into edge-based methods and density-based methods. The edge detection methods continue to detect the corner after finding the edges of the image as preprocessing. Artificial Neural Networks can be used to perform teaching the corner detection process. Artificial neural networks store and generalize information after the learning phase and relate the output produced by the system to the expected value. In this thesis study, an artificial corner catcher which learns corner information of objects by using Artificial Neural Networks for corner detection method in Scale Invariant Feature Transformation method is developed differently from classical approaches. The system was trained by creating positive and negative image parts and had high success results after test applications. Developed intelligent corner detection method was compared to other corner detection methods such as Moravec, Harris and Susan Corner Detection Methods and results were compared.