Image Blur Classification and Unintentional Blur Removal

Abstract

Blur is a general image degradation caused by low-quality cameras or intentional photographing for highlighting moving or salient objects. However, most blur classifiers just classify images into blur and sharp, which cannot distinguish the intentional blurred images from the unintentional blurred ones. Some unintentional blurred images are too valuable to discard directly. In this paper, we propose a robust image blur classifier to classify images into sharp, intentional blur, and unintentional blur. The basic idea of identifying the blur of a pixel being intentional or unintentional is that whether the blur occurs on a salient and semantic meaningful object. This inspired us to employ cues of blur, saliency, and semantic segmentation. We use spatial pyramid pooling to extract global features. Then, a random forest is used to conduct classification. We further detect the unintentional blur pixels by incorporating the cues into a conditional random field (CRF). The intentional blur image can be generated by pasting the deblurred unintentional blur regions back to the blur image. We conduct image blur classification on UBICD dataset and unintentional blur removal on different types of unintentional blur images. The experimental results show superior performance of image blur classification and the promising results of unintentional blur removal of our method.