HRC-Net : A High Robust ConvNeXt for Noisy Image
Chengfeng Hu, Jun Jie Zhu, Yang Yang · 2024
Despite the fairly good performance of Convolutional Neural Networks (CNNs) in image classification tasks, existing CNNs do not perform well when handling datasets with Gaussian noise. This results in the instability of model performance under extreme conditions or due to device aging. To address this issue, this paper introduces parameter-free attention and employs a simple wavelet decomposition and reconstruction method for image preprocessing, which minimizes the impact of noise on the model's performance. When the noise ratio is 40%, the accuracy of the improved model is approximately 90.4%, representing an increase of 3.1% compared to the original model's accuracy of 87.3%. This indicates that the proposed model in this paper is less sensitive to the impact of Gaussian noise, demonstrating its high robustness.