Accurate and Fast Blur Detection Using a Pyramid M-Shaped Deep Neural Network

Blur detection is aimed at estimating the probability of each pixel being blurred or non-blurred in an image affected by motion or defocus blur. This task has gained considerable attention due to its promising application fields in computer vision. Accurate differentiation of anomalous regions (including the sharp but homogeneous regions and pseudo-sharp backgrounds) and motion-blurred regions are main challenges in blur detection, in which both conventional and recently developed blur detection methods have limited performance and low time efficiency. To address these issues, this paper develops an accurate and fast blur detection method for both motion and defocus blur using a new end-to-end deep neural network. First, a novel multi-input multi-loss encoder-decoder network (M-shaped) is proposed to learn rich hierarchical representations related to blur. Then, to resolve the problem shows that blur degree is susceptible to scales, we construct a pyramid ensemble model (PM-Net) consisting of different scales of M-shaped subnets and a unified fusion layer. The experiments demonstrate that the proposed PM-Net can accurately handle those challenging scenarios with anomalous regions for both defocus and motion blur. Our method performs better than previous state-of-the-art methods. It achieves the $F_{1}$ -score of 0.893 for only defocus blur and 0.884 for joint motion and defocus blur, both of which significantly surpass previous methods on the benchmark BDD dataset. We also test our PM-Net on another public CDD dataset composed of challenging defocused images. The proposed method also outperforms other published methods with an $F_{1}$ -score of 0.885. In addition, our proposed method is hundreds of times faster (millisecond) than other state-of-the-art methods (second). Moreover, our experiments also demonstrate that the PM-Net is robust to noise and has a good generalization property.