A Coarse-to-fine Classification for Motion Blur Kernel Size Estimation with Cascaded Neural Networks

Blur kernel size estimation significantly affects the quality of image deblurring. However, existing methods have a large limitation in estimation accuracy as well as in the adaptability to large kernel sizes. In this paper, we propose a novel blur kernel size estimation method using CNNs. We first convert the 2-D kernel size estimation to a 1-D classification problem. A cascaded network model using homogeneous networks is then designed to cut the entire estimation space into 38 small intervals to have a more accurate kernel size estimation in a coarse-to-fine manner. Experiments show that the proposed method has reached a satisfying performance of kernel size estimation with a largely improved adaptability to various blur kernels.