Research On Image Super-resolution Reconstruction Technology Based On Mixture Model

Image super-resolution reconstruction aims to recover high-resolution images with clear details from blurred low-resolution images.In this process,the algorithm is able to recover lost details and sharpness,making the image sharper and more detailed.Image super-resolution is often used to improve image quality to better meet the needs of human vision and computer vision applications.In recent years,deep learning-based super-resolution reconstruction techniques have made breakthroughs in solving this problem.The performance of image superresolution reconstruction has been significantly improved and very good results have been achieved due to the large number of deep learning algorithms driving the reconstruction.The aim of this paper is to improve the existing convolutional neural network approach by proposing a hybrid model structure.The main research work of this paper is as follows:(1)To address the problem that CNN-based methods do not fully exploit the internal and external information of images,this paper proposes a lightweight Transformer-CNN hybrid model based on image super-resolution.A lightweight transformer(converter)structure is constructed to capture this information with internal and external interdependencies.At the same time,a dense block structure and residual connections are applied to construct a residual dense convolution block to reduce parameters and extract shallow features to some extent.The Lightweight Transformer Block(LTB)further extracts features and learns texture details between blocks through a self-attentive mechanism.EMT)with a small GPU memory footprint and benefits from multi-head attention,reduced and extended feature pre-processing.EMT significantly reduces the GPU resource usage.In addition,a detail-purifying attention block is proposed to explore contextual information in high-resolution space to recover more details.Experiments on four benchmark datasets validate the effectiveness of the proposed model in terms of quantitative metrics and visual effects.The proposed algorithm model occupies only about 40% of the GPU memory of other algorithms and has better performance.(2)To address the problem that the perceptual field keeps decreasing during the training of the network model and the extraction of features is limited,this paper introduces a hybrid null convolution joint group to solve this problem and proposes a super-resolution reconstruction model for images with two-channel attention based on null convolution.In this model,global jump connection and local connection are used to improve the extraction efficiency for both global and local information,while dense connection combined with residual blocks is used to further improve the mobility of feature information.In this model,a two-layer attention module with spatial and channel mixing is proposed for the correlation between information and channels at different locations in space,which helps the network to learn more detailed features and thus reconstruct images more accurately.It is experimentally verified that this model achieves superiority in both objective and subjective aspects on four benchmark datasets and has some advantages over other models.