Research On Image Super-resolution Reconstruction Technology Based On Deep Learning

Image super-resolution reconstruction is a research hotspot in the field of computer vision.It mainly refers to the use of a set of low-resolution images to generate a single high-resolution image.Image super-resolution reconstruction technology can overcome the shortcomings of hardware devices to obtain high-resolution images.It can improve image recognition capabilities and recognition accuracy and has broad development prospects.It is widely used in military,medical,public safety,face recognition and other fields.In recent years,image super-resolution reconstruction algorithms based on sparse coding and GAN have been widely used.This paper improves and innovates based on the two methods mentioned above.The main works are as follows:(1)This paper expounds the research background and significance in this field,summarizes the domestic and foreign research status in the field of image super-resolution reconstruction,and introduces the basic principles and basic theoretical knowledge of sparse coding and generating adversarial networks in detail.It proves the superiority of sparse coding and generative adversarial network in the field of image super-resolution reconstruction.(2)Aiming at the problem of poor scaling flexibility of the model and the ill-posedness of the reconstructed image,a sparse coding network-based image super-resolution model was proposed based on sparse coding.A learning-based fast iterative shrinkage threshold algorithm(LFISTA)is used to implement a feedforward neural network,and its layers strictly correspond to each step in the super-resolution reconstruction process of sparsely encoded images.At the same time,all components of sparse encoding can be Joint training through back propagation.Cascading multiple sparse coding networks makes the model suitable for arbitrary scaling factors and increases the flexibility of the model.(3)In order to obtain high-resolution images consistent with the human visual system,an image super-resolution reconstruction model based on conditional generative adversarial networks is proposed.Combining convolutional neural networks with generative adversarial networks to improve the generator network in conditional adversarial networks,and the feature images of different network levels are cascaded by adding a skip connection method.The discriminator network uses the traditional VGG16 network model.The experiments in this paper are tested on the Set5,Set14,Urban100,and BSD100 datasets,and compared with multiple super-resolution reconstruction algorithms.The experimental results show that the two reconstruction algorithms proposed in this paper have improved the reconstruction quality to a certain extent,and the reconstruction algorithms based on conditional generation adversarial networks can obtain better subjective visual effects.