Adversarial Learning for Image Super Resolution Using Auxiliary Texture Feature Attributes

Recently, image super resolution (SR) has made great progress in computer vision (CV) domain. Due to fine-grained features by convolution neural network (CNN), the mapping function among low resolution (LR) images and high resolution (HR) counterparts could be learned. Textures, which exerts relevant features measured by texture description algorithms, such as gray level co-occurrence matrix (GLCM), among local image patches, are such vital for image SR to recover high frequency details. However, most previous deep learning based methods directly estimate high resolution images ignoring these similar texture attributes. In this paper, we attempt to seek auxiliary texture feature attributes from image patches and cluster them into group indexes so as to improve the reconstruction quality. For the sake of more latent contextual information, a wide residual block is introduced. What's more, we employ 4 loss functions, including adversarial loss, perceptual loss, textural group-conditional loss as well as total variation (TV) loss, tending to estimate high resolution images with realistic textures during the adversarial training. Extensive experimental results on several benchmark datasets demonstrate the effectiveness of proposed method on improving visual quality.