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

Figure 6

Joint generative image deblurring aided by edge attention prior is used to locate the area which is red in the picture and deblur the patches aided with edge prior. During the process of deblurring, multiple blur kernels are adopted dynamically including motion blur and Gaussian blur and so on. (a) An original blur image. (b) Attention map of specific objects. (c) Receptive field of attention is large, and the scan is with large-scale objects. (f, g) The small receptive fields, respectively. The edge recovery of a generative adversarial network (GAN) tends to be slightly intermittent, but the restoration effect is the best in complex structures. (e–g) The experimental validation of the selective regions to deblur the specific categories of blur objects in multiscale.

(a)

(b)

(c)

(d)

(e)

(f)

(g)