Hossein Hosseinzadeh

Convolutional Neural Networks (CNNs) are widely used for image classification tasks. However, the presence of noise, especially impulse noise, can significantly reduce their performance. In this study, a noise-robust CNN architecture is proposed, where instead of employing complex denoising preprocessing techniques, a Median Pooling layer is utilized to enhance the model’s robustness against noise. The proposed model is evaluated on three datasets, including shoe images, fruit images, and traffic sign images, under three different scenarios: clean data, noisy data, and clean training with noisy testing. Experimental results demonstrate that the use of Median Pooling substantially improves the robustness of the model to noise, maintaining classification accuracy of noisy images close to that of clean data. Furthermore, the findings reveal that training solely on clean data leads to a considerable drop in accuracy under noisy conditions, while training on noisy data combined with Median Pooling provides the best balance between accuracy and noise robustness.