IMAGE SUPER-RESOLUTION USING COMPLEX DENSE BLOCK ON GENERATIVE ADVERSARIAL NETWORKS

• Main Authors: Bo-Xun Chen, 陳柏勛
• Other Authors: Tsung-Jung Liu
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/cgi-bin/gs32/gsweb.cgi/login?o=dnclcdr&s=id=%22107NCHU5441051%22.&searchmode=basic

碩士 === 國立中興大學 === 電機工程學系所 === 107 === The recent super-resolution (SR) techniques are divided into two directions. One is to improve PSNR and the other is to improve visual quality. Although convolutional neural networks (CNN) have recently demonstrated higher reconstruction quality in single-image super-resolution (SISR) than traditional methods, these methods typically minimize the generation of high-resolution images and actual high resolution. Pixel mean square error (MSE) or mean absolute error (MAE) or loss between degrees of image, so they produce a very high PSNR. Unfortunately, this approach averages possible solutions, minimizing these losses inevitably leads to edge blurring, so complex natural and realistic textures remain a challenging problem. We believe improving visual quality is more important and practical than blindly improving PSNR. In this paper we employ the network architecture using the concept of generative adversarial network (GAN) and a new perceptual loss function for photo-realistic SISR. Our main contributions are as follows: we propose a new dense block which uses complex connections between each layer to build a more powerful generator. Next, to improve the perceptual quality, we found a new set of feature maps to compute the perceptual loss, which would make the output image look more real and natural. Finally, we compare our results with other methods by subjective evaluation. The subjects rank the image generated by various methods from good to bad. The final results show that our method can generate a more natural and realistic SR image than other state-of-the-art methods.