#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ Created on Sat Jun 13 14:50:12 2020 @author: am2806 """ #img_name= 'Snow' f = open('HTML_Report_Paper_ID_10510.html','w') start = """

Training a Better Loss Function for Image Restoration

Anonymous ICCV 2021 Submission: Paper ID 10510

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.

""" f.write(start) f.write("

Applications

") f.write("

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.

") f.write("

Single Image Super-Resolution (SISR)

") f.write("Enhanced Deep Super-Resolution (EDSR)") f.write("
") f.write("Super-Resolution ResNet (SR-ResNet)") f.write("
") start1 = """

Additionally, we provide results for SISR on the General 100 Dataset. The images in this dataset are from different scenes and of different dimensions.

""" #f.write(start1) #f.write("

Noisy Image Reconstruction!

") #f.write("Enhanced Deep Super-Resolution (EDSR) on General 100 Dataset") #f.write("
") f.write("

Image denoising

") start2 = """

""" f.write(start2) f.write("Image denoising") f.write("
") f.write("

JPEG artefact removal

") start3 = """

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

""" f.write(start3) f.write("Compression qualiy = 7") f.write("
") f.write("Compression qualiy = 10") f.write("
") #Ablation Study #f.write("

Ablation Study

") #start3 = """ #

In this section, we provide comprehensive ablation results for our proposed Multi-Scale Discriminative Feature (MDF) loss function. #More specifically, we study how the performance of the loss function changes with the number of scales at which features are extracted, and the #effect of changing the seed image.

#""" #f.write(start3) # #f.write("Scales of Discriminators") #f.write("
") #f.write("Seed Image") #f.write("
") # #start3 = """ #

Further, we conducted experiments for SISR (EDSR model) #with different types of model initializations. The dataset used for training and testing the performance of the network is #Berkeley Segmentation Dataset (BSD)

#""" #f.write(start3) # #f.write("Different Model Initializations") #f.write("
") #f.write("Snow Image!") #f.write("
") #Alpha f.write("

Hyper-parameter tuning for VGG and LPIPS

") start3 = """

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.

""" f.write(start3) f.write("Hyper-parameter tuning") f.write("
") #Plots f.write("") f.close()