AAI_2025_Capstone_Chronicles_Combined

6 be using the Python library “IQA-PyTorch” (Chen & Mo, 2022), which includes a large collection of Image Quality Assessment algorithms ranging from more classical models to state of the art neural networks. Figure4 Distribution of Image Quality score for algorithms MUSIQ, NIMA and NIQUE

The quality distribution of algorithms “MUSIQ” (Ke et al., 2021) and “NIMA” (Talebi et al., 2018) give slightly higher score to the Fake images than the Real ones. Applying the Kolmogorov-Smirnov test in both distributions yields D = 0.166 and D = 0.1614 respectively with p < 0.001 in both cases, indicating that there is a statistically significant difference. This is not surprising, as generative models are usually trained to generate good quality images, but real life images may not be that perfect. The algorithm “NIQUE” (Mittal et al., 2012) though, seems to give roughly the same score. Background Information Several recent papers summarize the landscape of artificial image recognition, and highlight the importance of robustness and generalization of these methods (Altuncu et al., 2024). For starters, Convolutional Neural Network (CNNs) (Lecun et al., 1998) are the standard building block for most computer vision tasks. This architecture has been successfully tested in the past by Wang, S.-Y. et al. (2020), who

363