DR-Net: A Novel Generative Adversarial Network for Single Image Deraining

Figure 1

The architecture of the proposed DR-Net network. It takes a set of synthesized rainy images as inputs. First, these images are passed through a convolution layer. Then, two parallel convolution branches with five convolution layers are appended. After an additional layer and two convolution layers, the generator subnetwork outputs the derained image. Finally, the discriminator subnetwork uses the mixed derained images and the ground truths as inputs to distinguish if they are fakes or trues. Additionally, three skip connections between the front convolution layers and later convolution layers in generator subnetwork are added.

(a) Generator Network

(b) Discriminator Network