Training Task-Specific Image Reconstruction Loss

Anonymous WACV 2022 Submission: Paper ID 784

Due to size limitations, we include the first 30 images from each test set (out of 100). Images are stores as JPEGs with a quality 90 to ensure that coding distortions do not distort the results.

Applications

In this report we provide a comprehensive comparison of qualitative results for different loss functions across different applications. To begin with, we show results for two Single Image Super-Resolution (SISR) networks, namely, Enhanced Deep Super-Resolution (EDSR) and Super-Resolution ResNet (SR-ResNet). Further, we show the results for the applications of image denoising and JPEG artefact removal.

Single Image Super-Resolution (SISR)

Image denoising

JPEG artefact removal

We compare the performance of different losses for two codec compression qualities.

Hyper-parameter tuning for VGG and LPIPS

To find the best weightage, we conduct a hyper-parameter search over controlling the weightage sum of of VGG/LPIPS and MSE feature-wise loss fucntions: MSE + weight * VGG/LPIPS.