Image Super-Resolution Reconstruction Based On Semi-Coupled Deep Convolutional Sparse Coding Model

In modern society,images serve as one of the important channels for humans to obtain information.However,due to hardware limitations or requirements for transmission and storage,images may undergo degradation and become low-resolution images.In practice,it is often impossible to directly obtain the high-resolution images before degradation.Therefore,super-resolution reconstruction techniques are needed to enhance image quality.This paper is based on sparse coding theory and takes the essential relationship between images as the starting point.The primary focus is on image super-resolution reconstruction.Under the framework of deep learning,a semi-coupled deep convolutional sparse coding algorithm is proposed.The specific work is as follows:1.Previous deep unfolding models based on sparse coding or convolutional sparse coding mostly artificially set the prior term of the sparse coefficients as the ₁l norm.The artificial setting of the prior cannot effectively utilize the learning ability of deep neural networks and compromises the accuracy of the sparse coefficients.In this paper,deep neural networks are used to learn complex prior terms of sparse coefficients for both high-resolution and low-resolution images,effectively leveraging the learning capability of deep neural networks and ensuring the accuracy of the sparse coefficients.Additionally,the parameters of the degradation model are input into the network along with the images,providing more prior information to the network.Experimental results validate the effectiveness of the proposed algorithm,showing that the reconstructed images have richer texture details compared to other deep learning algorithms.2.Previous super-resolution reconstruction tasks based on sparse representation models usually assume strict equivalence or linear mapping relationships between the sparse coefficients of high-resolution and low-resolution images.However,linear or identity functions are too simplistic to accurately describe the complexity of the real world.Therefore,this paper improves the mapping relationship of sparse coefficients to a non-linear form and utilizes deep neural networks to learn a non-linear mapping function,pushing the mapping relationship to a more general level.The learning of the mapping function is integrated with the learning of sparse coefficient priors,considering the structural and relational aspects of the algorithm model.Experimental results demonstrate that the non-linear mapping function achieves better reconstruction performance in image super-resolution tasks.3.Considering that image super-resolution can be viewed as a transformation of different image forms within the same scene or target,there are similarities between photo-sketch synthesis tasks and super-resolution reconstruction tasks based on the nature of form transformation.To further explore the essential relationship between different forms of images,this paper directly applies the excellent reconstruction performance of the semi-coupled deep convolutional sparse coding algorithm-based deep unfolding model to the photo-sketch synthesis task without altering the network structure.This achieves an effective integration of image super-resolution reconstruction and photo-sketch synthesis models.Experimental results demonstrate the feasibility of applying this algorithm to photo-sketch synthesis tasks and show its advantages in image quality evaluation metrics compared to other style transfer algorithms.