High-frequency details enhancing DenseNet for super-resolution

Convolutional neural networks based models have made impressive advances for single-image super-resolution task. To advance the reconstruction quality of high-frequency details of the images, which are difficult to recover in super-resolution task, this paper proposes a super-resolution method using a high-frequency information enhancing densely connected convolutional neural network (SRDN) which can make the network pay more attention to high-frequency regions' reconstruction like edges and textures during training. Our method applies relatively higher weights on the gradient descent values of these high-frequency regions' pixels before they are propagated backward to update the parameters of the network during training. After that, we use a Generative Adversarial Network to finetune the trained model for finer texture details and more photo-realistic results. Experiments show that our approach can achieve a significant boost in the reconstruction quality of high-frequency details at high magnification ratios. We also design a novel measurement to evaluate the high-frequency details' difference (HFD) between the ground truth image and the generated image.