Weather Classification
When testing our weather classification model we obtained an accuracy of 96.7% without features and 91.6% when fine tuned with features. The accuracy of model without features surprisingly performed better than the model with features used. This can be because the features were added at the end in the linear layers. We aim to experiment more on feature addition in the earlier layers.
Image Restoration
The image restoration model was evaluated on both synthetic and real-world images. The metrics used for evaluation are PSNR and SSIM to compare the restored image with ground truth clear image. We obtained a validation PSNR of 25.41 and SSIM of 0.87 on the synthetic data and a validation PSNR of 17.23 and SSIM of 0.54 on the real world data. The low performance on the real world data can be attributed to domain shift. This can be overcome by either generating synthetic images more similar to real world or to replace the loss such that no ground truth would be required.
| Data | PSNR | SSIM |
| Synthetic data | 25.41 | 0.87 |
| Real-world data | 17.23 | 0.54 |
Qualitative restoration results
The trained restoration model is tested on fog images of varying densities. The images on the left are the degraded images given as input to the model and the images on the right are the output clear images obtained from the model. The model is robust to different fog densities and provides good restoration.
