The Application Of Deep Learning In Image Super-resolution

Image super-resolution,a high-resolution image generating from a low resolution image or image sequence,is an important subject in the field of computer vision.Deep learning uses convolutional neural networks to simulate the human brain for learning.It has many advantages,such as high level of abstraction and rich expression of features.Therefore,deep learning has strong applicability and advantages in the field of image super-resolution.This paper used the open source project as the basic model,and analyzed the code structure and model framework,optimized and improved on the basis of original code,then tested the result after optimization,the main work is as follows:Learning rate is an important parameter in deep learning,and it has a great influence on the speed and quality of training.If the learning rate is too small,it is easy to make the training efficiency of model be low and hard to weaken.If the learning rate is too large,it is easy to cause severe shock of loss value.In the process of model training,the learning rate needs to be adjusted constantly.This paper proposes an adaptive learning rate adjustment method,which dynamically adjusts the learning rate according to the results.The experimental results show the better performance of the model.The project model refers to the deep convolutional generative adversarial networks.Based on the analysis of its structure,this paper use the output of generative network as an incentive factor in loss function of adversarial network,which strengthen the link between the sub networks.Then we do some optimization for the two sub networks: increasing the number of standardization and improving activation function.The experimental results show that the optimization and improvement of the model can improve the overall performance.