FAMED-Ne：A Fast and Accurate Multi-Scale End-to-End Dehazing Network

The main contributions

1. It devise a novel multi-scale end-to-end dehazing net : learing priors of fast and accurate
2. The encoder-decoder architecture : a small model size

Compare with two model

The model is inspired by AOD-Net and FPCNet

1. In contrast to AOD-Net
   1. a fully point-wise CNN to regress K
2. In constrast to FPCNet
   1. a multi-scale encode
   2. a new training strategy

The proposed method

Our approach adopt the Gaussian\Laplacian pyramid architectures for mulit-scale fusion

1. The Single-Scale FAMED: FAMED-Net-SS
   

   1. There are five point-wise convolutional layers : the first four from the K-encoder and the last forms the decoder.
   2. The convolutional kernal of point-wise convolutional layers is $1\ast 1$ and the pooling kernal are different sizes.

2. The Multi-Scale Variants of FAMED-Net: FAMED-Net-GP and FAMED-Net-LP

3. FAMED-Net-GP

   1. The model down-sample the input hazy image to another two scales,1/2 scale and 1/4 scale.
   2. the estimated K : $K_{concat}=conv[K_1;K_2{|}_{\ast 2};K_3{|}_{\ast 4}]$
   3. Loss fucation : L2 loss
      

4. FAMED-Net-LP

   

   1. a\c is the same as above
   2. the estimated K: $K=K_s=K_{s+1}{|}_{\ast 2}+K_s^{\prime },s\in [1,2]$

5. About resize K and another

   1. Using bilinear interpolation to resize K
   2. Using fast-guided filter to refine the interpolation K , and the kernal of filter was set to 48
   
   

advantages and disadvantage

1. advantages ：
   1. The point-wise convolution plays a key role in constructing a compact and lightweight dehazing network
   2. Cascaded point-wise convolutional layers are very effective for tackling the dehazing problem by aggregating local statistic-based features layer by layer.
   3. The multi-scale architecture can handle scale variance in complex scenes and the low cost.
2. disadvntages:
   1. The transmission in the sky regions is incorrect, leading to under-dehazed artifacts (伪影)