Conclusions

Based on the different experiments we have performed, we can safely conclude that the model generalizes well for the below image restoration tasks.

• DeNoise

• JPEG Deblocking

• Super Resolution

• DeRain

• DeSnow

Application : We have demonstrated that using this model as a preprocessing block in the Detectron[7] Object Detection pipeline, we were able to detect more instances with higher confidence by removing noise from the images.

Generalization : The model generalizes well with other tasks like snow removal and rain drops removal.

Robustness : The model is also robust enough to different levels of attacks (noise, compression factor, scaling factor). The PSNR and SSIM values are higher for the recovered images when compared with the attacked images.

Extension : The model can be well extended to color images and the results have been demonstrated.

Ablation Studies :

• Increased the number of memory blocks to 7. This model has 7 memory blocks and each having 6 recursive units. (M7R6)

• Increased the number of filters used in our model to 128.

• Satisfactory and comparable results (PSNR, SSIM) obtained for both and the same has been demonstrated.