Impact of Gaussian Blur and Salt and Pepper Noise on CNN Performance using MNIST
Authors/Creators
Description
While Convolutional Neural Networks (CNNs) have achieved high accuracy in vision tasks, their robustness to common image distortions like Gaussian blur and salt and pepper noise is critical for real-world reliability.
This study quantifies the impact of these two specific corruption types on a simple CNN trained on the MNIST dataset. The CNN’s performance was systematically evaluated across controlled intensity levels of Gaussian blur and salt and pepper noise, isolating the effect of each degradation on classification accuracy. The results demonstrate the CNN is highly robust to smooth, low frequency disturbances (Gaussian blur), but critically fragile to sparse, high-contrast corruption (salt and pepper noise).This research highlights the necessity of noise-aware data augmentation and robust pre-processing filters for reliable de-ployment.