Frequency modularized neural network for deinterlacing

In this paper, a new model of the frequency modularized neural network for deinterlacing is proposed. In proposed method, image is divided into edge and flat regions by using its local frequency characteristic. And then, for each region, a neural network is assigned respectively. Since each region has similar pattern of information of the image, neural network can learn the similar patterns in frequency domain more easily. The input of neural network is ac component that is obtained by subtracting local mean from intensity of the pixel. It helps neural network to learn the input data more efficiently by removing redundancy due to the intensity of the pixel. In simulation, the proposed algorithm shows improved performance, compared with other algorithm and the method using the single neural network.