Image blur classification and blur usefulness assessment

Blur classification is an important and widely-studied problem in computer vision. State-of-the-art blur classification methods are designed and verified using man-made blur images of known blur types and blur kernels. In reality, natural blur occurs under wild-conditions, and thus cannot be simply simulated by several hand-crafted blur kernels. Hence, conventional blur classification methods cannot deal with complex real-world blur classification tasks. In this paper, we propose a new blur classification model, which learns from real-world images by convolutional neural network. On the basis of the blur classification network result, we further propose an interesting and useful problem, called blur usefulness assessment, which assesses the usefulness of blur image. To support blur classification and blur usefulness assessment, we establish a useful blur image classification dataset, UBICD, which contains 1; 000 sharp images and 1; 000 blur images (500 useful and 500 useless images). Compared with state-of-the-art blur classification methods, our method have achieved the highest blur classification accuracy of 98.1%. Our blur usefulness assessment also achieves an accuracy of 89.1%.