Improving surf interest point detection for defocus blur robustness

In this article, we propose a modification to SURF (Speeded Up Robust Features) to make the feature detection invariant to defocus blur. Specifically, SURF´s blob detection relies on the determinant of Hessian matrix constructed out of differential responses to the image. Our analysis of blur and its effect on SURF suggests that fourth derivative - and not the usual second derivative - is optimal for detecting the blurred blobs. The proposed defocus blur invariant SURF - which we refer to as DBI-SURF - does not require image deblurring nor blur kernel estimation, meaning that its accuracy does not depend on the quality of image deblurring.