Image Dehazing using Residual-Based Deep CNN

Implementation of paper https://ieeexplore.ieee.org/document/8355803 with the addition of the transmission refinement using guided filters from Dehazenet paper.

Dataset

Training Dataset

• NYU2 Depth Dataset - https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html

Used 1200 indoor images and their respective depth maps and generated 60000 image patches (16x16 size) 50 patch each image.

Testing Dataset

• Reside Standard Dataset - https://sites.google.com/view/reside-dehaze-datasets/reside-standard (Randomly taken indoor hazy images)
• Also taken few outdoor hazy images from Internet.

Training

Training is done on the dataset of size 16x16 in two phases:

• CNN Model - Hazy image is taken as input to generate the transmission map of the image.

  • Here we have also added the Transmission refinement using Guided Filter (taken from Dehazenet)

• Residual Model - Ratio of hazy image and transmission map is feeded to the network to generate the residual image.

• Residual Image is finally subtracted from the Hazy Image to generate the Clear Image.

Testing

Testing involves the creation of the network model from the image size and then loading the trained weights.

Network Testing Steps

1. Input Image (I)	2. Output Transmission map (t)	3. Transmission Map Refined (T)

4. Residual Map Input (H = I/T)	5. Residual Image (b)	6. Generated Clear Image (J)

Results

Outdoor

Input	Output

Indoor

Input	Output

For some images result were not as good

Problems Faced

This is my first attempt to implement a Deep Learning research Paper. Many things is not clear in the paper itself, so tried on the assumptions, like :

• The training dataset preparation is not specified properly, so used method mentioned in Dehazenet paper.
• Slice layer implementation is hard in Keras.
• Residual Network is not clear.

References

Datasets:

• https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html
• https://sites.google.com/view/reside-dehaze-datasets/reside-standard

Papers :

• DehazeNet: An End-to-End System for Single Image Haze Removal https://arxiv.org/abs/1601.07661
• AOD-Net: All-in-One Dehazing Network https://ieeexplore.ieee.org/document/8237773
• Recursive Deep Residual Learning for Single Image Dehazing https://ieeexplore.ieee.org/document/8575267/

Reference Codes :

• https://github.com/zlinker/DehazeNet
• https://github.com/yixindu1573/Recursive-Deep-Residual-Learning-for-Single-Image-Dehazing-DRL
• https://prateekvjoshi.com/2016/02/02/deep-learning-with-caffe-in-python-part-i-defining-a-layer/
• https://ddokkddokk.tistory.com/21
• https://stackoverflow.com/questions/42883547/intuitive-understanding-of-1d-2d-and-3d-convolutions-in-convolutional-neural-n
• keras-team/keras#890