LapSRN.md

Short introduction

Progressively reconstruct the sub-band residuals of high-resolution images.

Main contributions

• progressive training
• multi-prediction in one forward-pass

Architecture

Architecture

Loss

• Charbonnier penalty function
• Loss function:

Training strategy

Experiments

• Dataset: 91 images and BSD200

• Evaluation metric: PSNR and SSIM

• Patchsie: 128 × 128 × 64, augmentation including rotation and flip

• Adam, learning rate 0.00001, 50 epochs, 0.9momentum

• Results:

  Training with L2 loss generates SR results with more ringing artifacts.

  Intermediate predictions of our 8X model are slightly inferior to our 2X and 4X models.

  

Final summary

Pros:

• Good performance

Cons:

• No fine details on buildings. All SR algorithms failed except selfExSR. Large model.

Tips:

• Best results can be achieved by training with specific scale factors.