Convolutional Neural Network for Implementing a 2D Median Filter
Soonchul Jung, Jae Woo Kim, Yoon‐Seok Choi, Hyeong-Ju Jeon, Jin-Seo Kim · 2021 International Conference on Information and Communication Technology Convergence (ICTC) · 2021
While traditional neural networks could not solve the “exclusive or” (XOR) problem, it is known that current ones took the place of the universal Turing machine with the same level of computational capability. However, there was not enough vigorous research on how to construct neural networks to solve specific computational problems. Median filter is a nonlinear local signal processing technique that reduces noise from the original signal. It conducts more complicated operations compared to solving an XOR problem. In this paper, we developed a new neural network that performs 2D median filtering and conducted experiments by applying it to a variety of image data sets. Experimental results showed that the proposed XOR neural network could conduct median filtering with more than 99% accuracy. We observed that the accuracy is improved as the sizes of input images increase and also found that the accuracy was better for artificially created random images than natural images.