BPPNet : Back Projected Pyramid Network

       

Ayush Singh    Ajay Bhave    Dr. Dilip K Prasad

This is the official implementation of our ECCV 2020 Workshop paper Single image dehazing for a variety of haze scenarios using back projected pyramid network. Paper Link

Image dehazing refers to procedures that attempt to remove the haze amount in a hazy image and grant the degraded image an overall sharpened appearance to obtain a clearer visibility and smooth image. It has application in various fields such Outdoor Video surveillance, Driver assistance systems, etc.

We have implemented the network in pytorch

Experimental Details

Dataset used

We have trained and tested our 4 datasets named I-Haze, O-Haze, Dense haze and NTIRE2020 Non Homogenous Dehazing dataset

Training Details

The initial learning rate for both generator and discriminator was 0.001. The learning rate was decreased manually as the training proceed and the final learning rate for both generator and discriminator was 0.000001.

Results

The results of our model were quite good. As we can see from the below tables that our method clearly outperformed the state of the art with respect PSNR and SSIM both. Our model does the dehazing task in real-time at an average running time of 0.0311 s i.e. 31.1 ms per image.

I-HAZE

Metric	Input	CVPR’09	TIP’15	ECCV’16	CVPR’16	ICCV’17	CVPRW’18	Our model
SSIM	0.7302	0.7516	0.6065	0.7545	0.6537	0.7323	0.8705	0.8994
PSNR	13.80	14.43	12.24	15.22	14.12	13.98	22.53	22.56

O-HAZE

Metric	Input	CVPR’09	TIP’15	ECCV’16	CVPR’16	ICCV’17	CVPRW’18	Our model
SSIM	0.5907	0.6532	0.5965	0.6495	0.5849	0.5385	0.7205	0.8919
PSNR	13.56	16.78	16.08	17.56	15.98	15.03	24.24	24.27

Dense Haze

Metric	CVPR’09	Meng et. al	Fattal et. al	Cai et. al	Ancuti et. al	CVPR’16	ECCV’16	Morales et. al	Ours model
SSIM	0.398	0.352	0.326	0.374	0.306	0.358	0.369	0.569	0.613
PSNR	14.56	14.62	12.11	11.36	13.67	13.18	12.52	16.37	17.01

NTIRE2020 Non Homogenous Dehazing

As NTIRE2020 Non Homogenous Dehazing is very new dataset and its ground truth for validation and test data was not available, we were not able to compare it on the other state of the art. We randomly selected 5 images from training as testing images and trainined our model on the rest 40 images. The PSNR and SSIM for that testing data were 19.40 and 0.8726 respectively

Few of our results