Research On Image Super-resolution Algorithm Based On Deep Learning

Image is an important information medium.However,in the process of image acquisition,due to the limitations of the resolution and storage conditions of the acquisition device itself,the resulting images tend to have lower resolution.Image super-resolution technology is a kind of image post-processing technology.It can improve the resolution and visual effect of the original image only by algorithm processing without changing the hardware device.Therefore,this solution is low-cost,easy to promote,and has broad application prospects.Currently,image super-resolution algorithms can be divided into three categories: interpolation-based,reconstruction-based,and learning-based algorithms.The first two algorithms only treat the image as a signal,resulting in a poor visual effect on the processed image.The learning-based algorithm focuses on the image content itself.It uses existing training data to provide stronger priori constraints,and therefore can obtain better result images.At the same time,deep learning has a multi-nonlinear processing layer,which can extract low-to-high-level and specific-to-abstract features from the original data.Therefore,this paper will solve the image super-resolution problem through deep learning.The main research work of this paper has the following three points:(1)The convolutional neural network is introduced into the processing of image super-resolution problems.The main process of the algorithm can be divided into two phases of training and reconstruction.In the training phase,the original training image is first interpolated and scaled to obtain a down-sampled image,and then the original training image and the down-sampled image are divided into several image blocks,and finally constructed from these original image blocks and down-sampled image blocks.Training data set.By training on a data set,a model for low resolution images to high resolution images is obtained.For any given low-resolution image,the corresponding mapping relationship can be used to reconstruct the corresponding high-resolution image.Experiments show that the algorithm achieves better performance in objective evaluation.(2)An image super-resolution algorithm based on the introduction of generative adversarial network into the image super-resolution problem is introduced.The algorithm model is mainly composed of two parts: generator and discriminator.The generator receives the low resolution image as input and then outputs the simulated high resolution image.The discriminator accepts the simulated high resolution image and the original high resolution image and distinguishes them.The process of training is that the two confront each other and iteratively optimize.When the image output by the generator makes it impossible to distinguish the arbiter,the generator can be used to perform super-resolution reconstruction of any given low-resolution image.Experiments show that the algorithm achieves good performance both in visual effects and objective evaluation.(3)The training of the network model is based on minimizing the mean square error of the objective function.The result image obtained by this training method can only be consistent with the original image in the pixel,and lacks a good visual effect.Therefore,we introduce the perceptual error term to the objective function,shift the error calculation from the pixel space to the feature space,and use the training process to generate the discriminator constraint generator in the confrontation network so that it can produce a result image that is more in line with the visual experience.Experiment results show that the algorithm achieves good performance both in visual effects and objective evaluation.