AAI_2025_Capstone_Chronicles_Combined

16

Table4 Pearson’s test for relationship between AQI and misclassification rate for ViT TOPIQ_NR Fake r = -0.106 p<0.001 Real r = 0.021 p<0.001 To further explore the relationship between quality and accuracy, the effects of adding noise to correctly classified images were explored. Two approaches where tested: Gaussian noise and JPEG compression were applied to correctly classified images and the accuracy while varying the level of distortion was measured. Gaussian noise will add a per-pixel distortion that will interfere with short-range correlations. JPEG compression, on the other hand, affects the overall structure of the image and interferes with longer-range correlations. First, let’s begin by analyzing the performance of the CNN in these scenarios: Figure12 Accuracy relative to added Gaussian noise and JPEG compression, CNN

373