Kernel Functions in CNNs: Enhancing Robustness to Image Noise
Fengyu Gao, Hong Zhang, Hao Yan, Chien‐Hua Chen · 2025
Convolutional Neural Networks (CNNs) have achieved remarkable success across various fields, particularly in image processing. However, the interference from imaging devices and external environments introduces unavoidable noise, leading to a degradation in image quality. This paper proposes a Kernel-based Convolutional Neural Network (K-CNN) to enhance CNN robustness against image noise by incorporating kernel functions into convolution operations. The proposed K-CNN utilizes different types of kernel functions to improve nonlinear feature extraction without increasing model complexity. Experiments on the CIFAR-10 dataset demonstrate that K-CNN significantly outperforms standard CNNs in handling Salt-and-Pepper and Gaussian noise. The results highlight the potential of kernel-based convolution operations as an alternative approach for improving CNN robustness in real-world noisy environments.