In the second model, attentive GAN

In the second model, attentive GAN network is formulated to remove the rain impairment, its basic idea is to inject attention map to the original network. Given to the main idea mentioned above, a different rain image model is employed with attention map. The impaired images can still be viewed as the background images and rain mask, the rain mask can then be leveraged to guide the recurrent network generate rain attention map. The equation is expressed as follows, Where O is the input impaired images, M represents the attention map values from 0 to 1 , M=1 means the specific pixel belongs to the rain district, M=0 means the pixel does not belong to the rain district, B represents background image, S is the rain district, ʘ means element-wise multiplication. By leveraging the rain image model mentioned above, the whole task is equivalent to obtain a background image from a given input. Therefore, a raindrop attention map guided by a matrix M is created to better remove raindrops according to the raindrop area that the generator and discriminator should pay attention to. Since the raindrop area is mostly white, the corresponding grey value is large, so the rain image is first used to subtract the corresponding clean image, and then the threshold value of 30 and 15 is used to determine whether the pixel belongs to the rain area, the grey value would first divided by 30 and the pixel with value greater than 30 would be given to 1, and the pixel with value smaller than 15 would be set as 0 to generate a matrix M, and exploit M to guide the recurrent neural network to generate a raindrop attention map. By utilizing two thresholds, the generated attention mask would be considered as smooth in the rain area, as in real life the raindrop presents to show different appearance even in the rain area itself.