Performance Analysis of Oriented Feature Detectors

Abstract

Oriented feature detectors are fundamental tools in image understanding, as many images display relevant information in the form of oriented features. Several oriented feature detectors have been developed; some of the important families of oriented feature detectors are steerable filters and Gabor filters. In this work, a performance analysis is presented of the following oriented feature detectors: the Gaussian second-derivative steerable filter, the quadrature-pair Gaussian second-derivative steerable filter, the real Gabor filter, the complex Gabor filter, and a line operator that has been shown to outperform the Gaussian second-derivative steerable filter in the detection of linear structures in mammograms. The detectors are assessed in terms of their capability to detect the presence of oriented features, as well as their accuracy in the estimation of the angle of the oriented features present in the image. It is shown that the Gabor filters yield the best detection performance and angular accuracy, whereas the steerable filters have the best performance in terms of computational speed.