Joint Generative Image Deblurring Aided by Edge Attention Prior and Dynamic Kernel Selection

Figure 1

Various deblurring network architectures. (a) Nah et al. [13] proposed the multiscale architecture to extract features from different scales. (b) Tao et al. [15] proposed the recurrent architecture, in which the next round of training can be aided by the last round results. (c) Zhang et al. [17] utilized the multipatch architecture to directly extract features from image pairs by cropping images in different scales. (d) Ye et al. [18] used the scale-iterative architecture to train the model with an upsampling path with aid of the last-iterative middle results. We combine the ideas of (a) and (b) and propose a new framework whose core module involves the MRF and call it MRFNet. The MRFNet can operate in both multiscale and recurrent manner.