Corner Detection In Color Images Through A Multiscale Gabor Filters

Corners are important image features,and it plays an important role in image feature extraction.Color image contains more information and the variety information make image processing more complex.So a high-accuracy and real-time processing corner detector is very important.This paper proposes a contour-based corner detector for color images,which overcomes the problems in extraction of corner information using two kinds of corner detection algorithm for color images including Harris-based and contour-based.Firstly,make color images into grayscale using single channel method.Then in order to get a complete contour,it extracts the edge map of image using the classic Canny edge detector and fill the gap caused by the operation of non-maxima suppression using edge tracking method.Next,a set of Gabor filters consisting of five scales and eight directions are constructed to smooth the contour.For each contour pixel,the normalized sum is calculated by summing the magnitude response of the imaginary parts of Gabor filters of image in all direction at certain scale.Finally the new corner measure is specified as the product of the normalized sums at all scale.When the corner measure is not only above a previously specified threshold but also the local maximum,this contour pixel will be labeled as a corner.This evaluation and analysis technique uses two metrics average repeatability and localization error together.Compared with three other color corner detectors,the result from the experiment show that the proposed detector is more competitive with respect to detection accuracy,localization accuracy,affine transforms and noise-robustness.