Research On Image Super Resolution Algorithm In Natural Scene

People often use the shooting function of mobile phones or downloading and accessing pictures from the Internet.Due to the limitation of the function of the camera equipment and the external conditions such as the light in the shooting environment cannot meet the requirements,access devices such as mobile phones or computers are also A large number of low-resolution images have been accumulated,so it is of great significance to use software methods such as combining deep learning to increase the size of low-resolution photos.In recent years,with the development of deep learning theory,a large number of excellent super-resolution algorithms have emerged.However,the low-resolution image super-resolution in the real scene will be affected by many factors such as noise,artifacts,and unknown down-sampling kernels,and the improvement effect is still not ideal.This article conducts in-depth research on the above issues.The main content and results include:(1)A low-resolution image generation algorithm based on image frequency separation is constructed.Use the generative adversarial network to learn the data distribution of low-resolution images,and use the generator to simulate the output of the down-sampling kernel as the negative sample part of the new dataset.The filter divides image into high-frequency and low-frequency parts,which simplifies the training difficulty of the discriminator,and adds data enhancement processing.The experimental results prove that the generation network can learn the data distribution of real scene pictures,and training on the real dataset,the super-resolution model has improved the processing ability of real scene pictures.(2)Propose a super-resolution algorithm based on recurrent network.By saving the dense features extracted by the residual block,the output of the fusion residual module after each iteration is retained,so that the model has a memory similar to a cyclic network.A large number of fusion operations are used in the model.With the fusion of the local feature matrix and the global feature matrix,the receptive field of the model is improved and the complexity of the model is reduced.Improved the L1 loss function commonly used in super-resolution algorithms.When the difference between the generated image and the real image is less than the preset constant ?,this part is discarded from the calculation of the loss function,which relaxes the constraints on the generation ability of the network.The experimental results show that,compared with the commonly used super-resolution algorithms,the super-resolution algorithm proposed in this paper has obvious advantages for the restoration of low-resolution images of real scenes.