Human-Designed Filters May Outperform Machine-Learned Filters

Machine-learned image processing systems in medical imaging have shown better results than those obtained by traditional human-designed techniques. The success of machine learning techniques inspires humans to design better systems. The convolutional neural network (CNN) has a multi-channel architecture, which the conventional filters do not have. This paper proposes that by borrowing the multi-channel architecture, the human-designed denoising filter can have better performance than the machined-learned version. We illustrate the feasibility of this idea with a toy example in a sinogram denoising task in the area of tomography.