Research On Iterative Back-projection Image Super-Resolution Reconstruction Based On Attention

Image super-resolution reconstruction as a low-level vision task aims to extract information from low-resolution images and reconstruct high-resolution images.With the continuous development of deep learning,deep neural networks have gradually become the mainstream model for image super-resolution reconstruction,among which convolutional neural networks are the most widely used.Recently,thanks to the self-attention mechanism inherent in Transformers,visual Transformers have performed excellently in advanced computer vision tasks,and their performance in low-level visual tasks is also highly anticipated.Against this background,this thesis continues to delve into convolutional neural network-based superresolution reconstruction algorithms in order to further improve reconstruction performance.Additionally,it explores the application of visual Transformers and their lightweight versions in super-resolution reconstruction tasks,expanding the research scope.The main research contents are as follows:(1)In response to the single-image super-resolution reconstruction problem,this thesis proposes a Holistic Attention Back Projection(HABP)network based on global attention mechanism.The network consists of up-down iterative projection modules and a holistic attention module,which work together to capture more feature information.The model captures the correlation between different resolution features by converting between the up-down projection modules,which helps to learn high-frequency texture details and strengthen the feedback information transfer between modules.Attention mechanisms have been proven to effectively preserve rich feature information at each feature layer.In this thesis,layer attention mechanism is used to weight the feature layers outputted by the up-down projection,in order to find interdependent relationships between different layers,channels and positions.In addition,spatial channel attention mechanism is also used to guide the network to learn related information within and between channels.Through the collaborative action of the two attention mechanisms,both global and local information are effectively utilized.(2)To explore the application of visual Transformers in image super-resolution tasks,this thesis proposes an Improved Transformer Image Super-Resolution(ITSR)reconstruction network.The model introduces a unique multi-head transposed attention mechanism and a feedforward mechanism to improve the reconstruction effect of texture details,and ensures high performance when processing images of various scales.Additionally,given the successful performance of the up-down iteration approach in HABP,this idea is also incorporated into ITSR to reduce computational costs and maintain efficient performance.(3)Based on the ITSR network,the Improved Transformer Lightweight Network(ITLN)was proposed with a focus on model compression.Compared to ITSR,ITLN significantly reduces the number of model parameters while ensuring the same level of performance,with only 23% of ITSR’s parameters.ITLN replaces the computationally expensive sampling method used in ITSR and appropriately reduces the number of network layers.In addition,ITLN incorporates a feature complexity processing mechanism,where different complexity levels of image blocks are directed through different branches of the network.This mechanism not only effectively reduces the number of model parameters and enhances operating efficiency,but also adjusts the network optimization approach.