1. Dense Scene Information Estimation Network for Dehazing
(Dense Scene Information Estimation Network for Dehazing)
Decoder.A: Estimated ambient light A
Decoder.t: Estimated transmission map t
Decoder.J: directly restore the defogged image, provide guidelines for Decoder.A and Decoder.t during the training process, so that the restored image is closer to the real image Ground Truth, and all its outputs will be connected together and input to refine10- 13 in
2.BidNet: Binocular image dehazing without parallax estimation (CVPR2020)
Figure (a) The overall architecture of the binocular image dehazing network (BidNet). BidNet inputs binocular blurred image pairs and outputs fog-free binocular image pairs. (b) Structure of the transformation module (STM). STM is proposed to explore and encode the relationship between binocular image pairs
As shown in the figure, the BidNet network inputs the binocular foggy image pair and estimates the transmission map and atmospheric light, and at the same time dehazes the binocular image pair. The architecture of BidNet is shown in Figure (a). The STM (Stereo Transformation Module) is introduced to explore and encode the correlation between the binocular left and right images.
The overall structure is divided into two parts, one is the Stereo Transmission Map Estimation Network (TMENet Stereo Transmission Map Estimation Network) part, and the other is the Atmospheric Light Estimation Network (ALENet Atmospheric Light Estimation Network) two parts. The first part is the first part, which is the main part of the network, which is divided into three small parts, the feature extraction part of weight sharing, the stereo transformation module STM, and the refinement module. The input of the feature extraction part is Il, which is the left-eye image and Ir, which is the right-eye image, and they pass through an encoder-decoder network with shared weights (Figure b). After the output of the two enters the STM module, the left perspective tl and the right perspective tr are obtained, followed by the specific structure of STM, which borrows from the structure of non local networks. The last part is the refinement part, using simple average pooling of several layers to give it a little scale information, and then get the final tl and tr, and then use the standard formula to get the dehaze map Jr and Jl. Then add the base layer convolution to refine the final result and complete it. After that is the second part of ALENet, which uses Il as input and gets A(x) after passing through an encoder-decoder structure.
3.Densely Connected Pyramid Dehazing Network
(Milian Pyramid Defogging Network DCPDN)
In this paper, we propose a new end-to-end single image dehazing method - Densely Connected Pyramid Dehazing Network (dcpdn), which can jointly learn the transmission map, atmospheric light and dehazing. End-to-end learning is achieved by directly embedding the atmospheric scattering model into the network, thus ensuring that the method strictly follows the physics-driven scattering model. Based on the idea that dense networks can maximize the information flow of features at different levels, a new edge-preserving densely connected encoder-decoder structure based on multi-level pyramid pooling modules is proposed for estimating the transfer map. The network is optimized with a newly introduced edge-preserving loss function. To further fuse the mutual structural information between the estimated transfer map and the deblurred result, we propose a joint discriminator based on the generative adversarial network framework to judge whether the corresponding deblurred image and estimated transfer map are real or fake. Through ablation experiments, the effectiveness of each module in estimating transmission maps and deblurring results is verified. Extensive experiments show that the proposed method achieves significant improvements over existing methods.
Transmission Graph Estimation (A Survey of Proposed Pyramidal Densely Connected Transmission Graph Estimation Networks): A densely connected encoder structure is proposed that exploits the features of CNN's multi-layers, where dense blocks are used as the basic structure. The reason for using dense blocks is that it maximizes the information flow along these features and ensures better convergence by connecting all layers. Furthermore, a multi-level pyramid pooling module is employed to refine the learned features by taking "global" structural information into account in the optimization. To utilize the predefined weights of dense networks, we adopt the first Conv layer and the first three Dense blocks, and the corresponding downsampling operation transition blocks from the pre-trained dense network121 as the encoder structure. The feature size at the end of the encoding part is 1/32 of the input size. In order to reconstruct the transmission map to the original resolution, five dense blocks with refined up-sampling Transition-Blocks are stacked, and the feature maps of the same dimension will be connected in series. The reason for the 1/4, 1/8, 1/16, and 1/32 multi-level pooling at the encoding part: to obtain more global information, and to preserve multi-scale information as much as possible when estimating the transmission map.