Efface-the-haze

The main aim of this project is to apply CV techniques and develop pipeline for image enhancement via Dehazing. This project is carried out as part of Computer Vision course CSL7360 under the guidance of Dr. Mayank Vatsa.

Experimental Setup

Images are resized to 128x128 for the below experiments

Dark channel prior (DCP) for single image haze removal was implemented. Along with this the preprocessing technique such as White Balance (WB) and postprocessing techniques CLAHE(Contrast Limited Adaptive Histogram Equalization) and DWT(Discrete wavelet Transform) were also implemented.
Inference on RESIDE test dataset (SOTS Indoor & SOTS Outdoor) was carried out on FFA-Net pretrained model.
Involuted U-Net architecture with custom loss is created and trained on RESIDE dataset. Evaluation is carried out using RESIDE test dataset (SOTS Indoor & SOTS Outdoor).

Custom loss is defined as below :

Composite loss = 0.6 * Perpetual Loss (AlexNet) + 0.1 * SSIM Loss + 0.3 * PSNR Loss

Datasets and Architectures

Dataset	Architecture	Description
RESIDE	FFA-Net	FFA-Net: Feature Fusion Attention Network for Single Image Dehazing (AAAI 2020)
RESIDE	Involuted U-Net	U-Net architecture augmented with Involution

FFA-Net: Feature Fusion Attention Network

Training

Training is carried out Training is carried out for 15 epochs with SGD optimizer, using learning rate 1e-2, weight decay of 0.01 and momentum 0.9 on Involuted U-Net

Results (128x128 Size)

Methods	Indoor	Indoor	Outdoor	Outdoor
	PSNR	SSIM	PSNR	SSIM
DCP	14.77	0.7757	22.65	0.9226
DCP with Preprocessing and Postprocessing(Pipeline1)	11.83	0.678	15.7	0.7712
DCP with Preprocessing and Postprocessing(Pipeline2)	11.15	0.5116	14.93	0.6395
FFA-Net on pretrained model	14.86	0.5559	19.41	0.6363
Ours (Involuted U-Net)	16.85	0.6073	17.78	0.5790