Image-Super-Resolution

This repository implements a novel Knowledge Distillation based Generative Adversarial Network (KD-GAN) approach for achieving state-of-the-art image super-resolution. The proposed method surpasses existing techniques like SRCNN and SRGAN in terms of image quality, reaching a Structural Similarity Index Measure (SSIM) of 94%.

Key Features:

• Superior Image Quality: Achieves exceptional SSIM scores, outperforming existing methods.
• Hybrid Deep Learning Approach: Combines knowledge distillation, a refined loss function, and regularization techniques.
• Faster Training: Improved loss function facilitates faster training compared to traditional methods.

This repository includes:

1. Implementation of the KD-GAN model for image super-resolution.
2. Training scripts and configurations.

Getting Started:

1. Clone the repository: git clone https://github.com/Ghaayathri-Devi-K/Image-Super-Resolution
2. Install dependencies (Refer to requirements.txt for details).
3. Utilize the provided scripts for image super-resolution using the trained model.

Experiment

The model was trained using Python v3.9.13, and TensorFlow library v2.11.0 in a local machine configuration with 16GB RAM and NVIDIA GeForce RTX 3060 GPU.

Dataset Used:

• Set5: It contains five high-quality images, offering a range of different textures and structures.
• Set14: Comprising 14 images, this dataset provides a more comprehensive challenge with a variety of scenes and objects.
• URBAN100: Focused on urban scenes, URBAN100 contains 100 high-resolution images. It is useful for evaluating how well super-resolution models handle man-made structures.
• BSD100: Part of the larger Berkeley Segmentation Dataset, BSD100 includes 100 diverse natural images.

Model Architecture Used:

1. Backpropogation steps used in the training process of the proposed KD-GAN

The model architecture described in this section was trained for 100 epochs with a learning rate of 0.005.

2. The generator

Accepts an input of any size but requires a depth of 3 channels (RGB).

3. The discriminator

Results

1. Results obtained from mathematical models and Random forest regression and the original image displayed for comparison

2. Results from image super-resolution using SRCNN

3. Results from image super-resolution using KD-GAN

4. Comparison of different models for the test-dataset

4. Comparison of different models for the existing benchmark image super-resolution datasets