This work used computational resources from the Center for Computation and Data Analysis (CCAD) of the National University of Córdoba (UNC), which is part of the National System of Computation and Data Analysis (SNCAD) under the Ministry of Science and Technology of the Argentine Republic (MinCyT).
This repo contains the most relevant codes from my PhD thesis.
The goal of this project is to compare Image Quality Assessment (IQA) similarity indexes by leveraging them as loss functions for Deep Neural Networks (DNNs) that perform low-level vision tasks [1]. The DNN architectures are defined in the DDN.py file, while the loss functions based on IQA models are implemented in the IQA_models.py file.
The DNN architectures used in this project are based on the EDSR (Enhanced Deep Super-Resolution Network) architecture presented in the work of Lim et al.[2].
Figure: DNN architecture based on the EDSR model
The IQA models used in this project are derived from the following sources:
- IQA-PyTorch, a collection of image quality metrics implemented in PyTorch, including both traditional and deep learning-based metrics.
- PIQ (Photo-Image Quality), more details can be found in the documentation.
For the training we used the Waterloo Exploration Database [3], which currently contains 4744 pristine natural images covering a variety of subjects. We duplicated this dataset by contaminating all images with Gaussian noise (add_noise_to_pristine_images.py), as the DDNs learn from pair of images noisy-clean.
In the following folders:
- Quality Analysis
- Spatial Statistics
you will find various analysis methods that can be applied to the denoised test images (get_denoised_images.py). These images were denoised by the trained DNNs and were not used during the training phase. By analyzing these results, we can assess the performance of each DNN and, ultimately, compare the effectiveness of the IQA indices through their corresponding evaluation.
- Ding, K., Ma, K., Wang, S. & Simoncelli, P. (2021). Comparison of Full-Reference Image Quality Models for Optimization of Image Processing Systems. International Journal of Computer Vision, 129, 1258–1281. Link to paper.
- Lim, B., Son, S., Kim, H., Nah, S. & Lee, K. M. (2017). Enhanced Deep Super-Resolution Network. IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 1132-1140. Link to paper.
- Ma, K., Duanmu, Z., Yong, H. & Wang, Z. (2017). Waterloo Exploration Database: New Challenges for Image Quality Assessment Model. IEEE Transactions on Image Processing, 26(2), 1004-1016. Link to database.