Defocus blur radius classification using random forest classifier

Abstract

Image blurring is an issue leading to degradation of image and thereby loss of information. An image is blurred due to motion between the camera or sensor and the object, defocusing of the lens, atmospheric turbulence, etc. Images can be restored if information about the original image or the blur is known. Degradation due defocus blur is considered in this paper. The estimation of defocus blur radius is treated as a classification problem here. A technique to classify the blur radius using random forest classifier is proposed in this paper. Third and fourth statistical moments, calculated on the spectrum of the blurred image are used as features for classification in the proposed method. The proposed technique is applied on images from the USC-SIPI Image database. The experiment results display defocus radius classification accuracy of 94.37%. The selection of skewness and kurtosis as handcrafted features for the defocused blur image and usage of random forest classifier for defocus radius classification has been approached for the first time in such problem. The choice of features adds to the novelty and efficiency of this approach.