The main contributions
- It devise a novel multi-scale end-to-end dehazing net : learing priors of fast and accurate
- The encoder-decoder architecture : a small model size
Compare with two model
The model is inspired by AOD-Net and FPCNet
- In contrast to AOD-Net
- a fully point-wise CNN to regress K
- In constrast to FPCNet
- a multi-scale encode
- a new training strategy
The proposed method
Our approach adopt the Gaussian\Laplacian pyramid architectures for mulit-scale fusion
-
The Single-Scale FAMED: FAMED-Net-SS
- There are five point-wise convolutional layers : the first four from the K-encoder and the last forms the decoder.
- The convolutional kernal of point-wise convolutional layers is 1 ∗ 1 1*1 1∗1 and the pooling kernal are different sizes.
-
The Multi-Scale Variants of FAMED-Net: FAMED-Net-GP and FAMED-Net-LP
-
FAMED-Net-GP
- The model down-sample the input hazy image to another two scales,1/2 scale and 1/4 scale.
- the estimated K : K c o n c a t = c o n v [ K 1 ; K 2 ∣ ∗ 2 ; K 3 ∣ ∗ 4 ] K_{concat}=conv[K_1;K_2|_{*2};K_3|_{*4}] Kconcat=conv[K1;K2∣∗2;K3∣∗4]
- Loss fucation : L2 loss
-
FAMED-Net-LP
- a\c is the same as above
- the estimated K: K = K s = K s + 1 ∣ ∗ 2 + K s ′ , s ∈ [ 1 , 2 ] K=K_s=K_{s+1}|_{*2}+K_s' ,s∈[1,2] K=Ks=Ks+1∣∗2+Ks′,s∈[1,2]
-
About resize K and another
- Using bilinear interpolation to resize K
- Using fast-guided filter to refine the interpolation K , and the kernal of filter was set to 48
advantages and disadvantage
- advantages :
- The point-wise convolution plays a key role in constructing a compact and lightweight dehazing network
- Cascaded point-wise convolutional layers are very effective for tackling the dehazing problem by aggregating local statistic-based features layer by layer.
- The multi-scale architecture can handle scale variance in complex scenes and the low cost.
- disadvntages:
- The transmission in the sky regions is incorrect, leading to under-dehazed artifacts (伪影)