TLQS Deep Learning S/W

This repository contains the S/W framework used for Deep Learning-based image super-resolution experiments in TLQS research work. The corresponding paper will be opened and linked here once accepted.

Software Requirements

• tensorflow2 (<= 2.15.1)
• python3
• matplotlib
• tqdm

Instructions

Training

1. Set hyper-parameters properly in main.py such as batch size and learning rate.
2. Run training as follows.

python3 main.py

Output

This program evaluates the trained model after every epoch and then outputs the results as follows. The super-resolution performance is measured as PSNR db.

Epoch: 0  learning_rate: 0.0010000000474974513  Bicubic_loss: 0.002334445716184646  train loss: 0.002353070449214034
psnr_bicubic_mean: 31.771656417846682  psnr_output_mean: 32.056907653808594  diff: 0.28525123596191193

Results

The experimental results will be available in the paper once it is published.

Supported Training Algorithm

• Mini-batch Stochastic Gradient Descent (SGD)

Datasets

• VDSR images / Set5 (link)
• Random sequences for shuffling data are included as 'npy' file in the repository.

Questions / Comments

• Jihyun Lim ([email protected])
• Sunwoo Lee ([email protected])