AAI_2025_Capstone_Chronicles_Combined

14 Next, misclassified samples were analyzed using AQI algorithms to look for any obvious correlation. False positives and false negatives where grouped into bins according to their quality values, and they were normalized to show the misclassification density relative to the total number of samples in the bin. For the CNN results, algorithms “PIQE” (N et al., 2015) and “WADIQAM_NR” (Bosse et al., 2017) stand out as they clearly show some clear correlation. Figure10 Relative misclassification rate compared to different AQI metrics for CNN

PIQE is an AQI algorithm based on classical algorithms that assigns a score from 0 to 100, where lower means better image quality. WADIQAM_NR, on the other hand, is a neural network based approach where higher score means better image quality. Applying Pearson’s test to the relationship between the AQI scores and the relative error rate shows a significant association between image quality and misclassification. For both algorithms, misclassified Fake images have substantially worse quality than correctly classified Fakes, while misclassified Real images have substantially higher

371