Results

We have used the PASCAL VOC2007 dataset for training De-noise, JPEG Deblocking, Super Resolution and combined model. We used the PASCAL VOC2007 for creating the synthetic datasets for each of the image restoration use-cases.

We have trained 6 different models for 100K iterations for the purpose of this project.

1. De-noise model - Generated images with noise level - 0.1 and trained the model for this dataset. Salt and pepper noise is applied on ground truth image to generate noisy images.

2. JPEG Deblocking - Generated images with compression factor - 10 and trained the model for this dataset.

3. Super-Resolution - Generated images with scaling factor - 1/3 and trained the model for this dataset. Dataset is generated by downscaling the ground truth image and then upscaling by 3x to generate train image. Nearest neighbor interpolation is used to scale the image. Goal of this model is to increase the clarity of the output image.

4. Combined Model - Performed data augmentation on images using noise level 0.1, compression quality 10, scaling factor 1/3 and trained the model on this synthetic dataset.

5. DeRain - We have used the DeRain dataset and extended the existing MemNet to support training of color images for DeRain task.

6. DeSnow - We have used the DeSnow dataset and extended the existing MemNet to support training of color images for DeSnow task.

Performance Evaluation Metrics used:

1. Average PSNR

2. Average SSIM

3. Training Loss

Plots for the above have been plotted and can be seen in the Performance Evaluation section.