Area-Specific Convolutional Neural Network with Multi-Path Method

Recently, with the development of artificial neural networks, deep convolutional neural network (CNN) has been successfully introduced in the single image super-resolution (SISR) field. Among several CNN-based image restoration methods, area-specific convolutional neural network (ASCNN) is proposed based on the observation that low-frequency pixel areas have relatively less influence on image restoration quality than high-frequency pixel areas. After ASCNN analyzes the spatial frequencies of an image, it applies a relatively simpler convolution operation to the low-frequency region than to the high-frequency region. However, this method cannot perform optimized operations on each pixel. In this paper, multi-path method is proposed which further subdivides the image area into more than two areas. By performing operations only as needed according to the frequency level of each pixel, FLOPs (floating-point operations) can be efficiently reduced even with a slight decrease in performance. According to experimental results, instead of a decrease of approximately 0.43% in PSNR (peak-topeak signal ratio), FLOPs is reduced by up to 42%.