MATLAB implementation of ECCV 2018 paper "Proximal Dehaze-Net: A Prior Learning-Based Deep Network for Single Image Dehazing" with several improvements. We provide the full functions of training and evaluation as well as our trained network models.
This work is implemented based on MatConvNet package and we have included Linux pre-compiled files in this project. To train our network, we require a good GPU device and CUDA toolkit.
We provide several trained models for directly network evaluation. The models are located in ./models/. For an example of using these models for image dehazing as shown in :
% use network with one stage trained on our dataset
% image: input hazy image
% resim: recovered haze-free image
% restt: estimated transmission
[resim, restt] = ours_tiphqs_s1_eval(image)
% use network with two stage trained on RESIDE dataset
% image: input hazy image
% resim: recovered haze-free image
% restt: estimated transmission
[resim, restt] = ours_tipres_s2_eval(image)To train our network, the training dataset must be generated first. We offer the training images and code to generate training data. Please download from Baiduyun. After generating training data, move them to ./data/train/ or modify opts.imdbPath in each training file to the data path.
Run following code will then train all networks:
run_train