AAI_2025_Capstone_Chronicles_Combined

17

While the model was fairly resistant to JPEG compression artifacts, when applying Gaussian noise its accuracy on Real images quickly drop, while the accuracy on Fake images keeps roughly the same. Figure13 Examples of altered images, CNN

This seems to show that the accuracy of the model is highly dependent on the noise of the image. To verify this, correctly classified images were “denoised” by applying a blur filter with different radii. Figure14 Accuracy vs blur and image example, CNN

374