Idea:We model the dehazing problem as a Markov Decision Process (MDP) with several existing simple traditional image processing operations and prior knowledge-based dehazing methods as actions
We use a deep reinforcement learning framework Deep Q-Network(DQN)
Agent Network
-
The role of Agent Network : Exteract Future.
-
The structure of Agent Network : The agent network consists of four fully connected layers,and the first three layers are followed by a ReLu layer.
-
The effect of Agent Network : It can effectively capture the inside cue ,and promote the seclection of correct action during the descion sequence.
-
The updata of Agent Network : Via the Q-learning process by Eq.(2)
1. Q\*(s,a) = real time reward r + optimal action a' selected at the next moment in the Q*(s',a')
-
The loss fuction of Agent Network
-
The loss fuction of agent network is defined as
-
The gradient of loss fuction is defined as
-
Action space A
- Add a D C P ∗ = m i n ( D C P ( x ) , x ) DCP^* = min(DCP(x),x) DCP∗=min(DCP(x),x) to improve the performance in the white areas of the image.
- Define Actions of 5 to 10,which transform color channel’s value , and action 11 represents that theimage dehazing processing should be done .
reword function R t R_t Rt
-
Adopt both PSNR and SSIM metrics as the measurement
-
Based on L2,PSNR and SSIM , the artical design the reward fuction by :
Farmework reinforcement learing
Classification of reinforcement learing :
- value based -> DQN
- poli based
- actor critic