초해상도를 위한 Depthwise 컨볼루션 활용 영역별 컨볼루션 신경망

Abstract

Super-resolution(SR) is the image processing that reconstructs high-resolution images from low- resolution inputs. Representative models such as ASCNN distinguish between high and low frequency regions and apply a Low Parameter Convolution(LPC) with reduced channel dimensions to the low frequency regions. Although various other lightweight models have also been proposed, balancing computational cost and reconstruction quality remains a significant challenge. In this paper, we propose DW-ASCNN, a lightweight variant of the original ASCNN, in which the LPC layers are replaced with depthwise and pointwise convolutions. The proposed structure identifies distinct convolutional paths by leveraging the frequency characteristics of each region in ASCNN, and achieves parameter and computational efficiency through structural simplification of these separated paths. Experimental results show that DW-ASCNN reduces the number of parameters by approximately 24% and lowers FLOPs by up to 7.93% compared to the original ASCNN, while keeping the PSNR degradation within an average of 0.02 dB. These results show that the proposed model is well-suited for deployment in resource-constrained environments.