Research On Image Super-resolution Reconstruction Based On Deep Learning

With the rapid development of computer science and technology,there are more and more images appeared in our daily life,and the requirements of image quality are also increasing.Image resolution is an important indicator to describe image quality.The techniques for improving image resolution mainly include hardware technology and software technology.However,improving the precision of hardware devices is costly and technically difficult.Therefore,improving the resolution of images through software technology has become a research hotspot.Algorithms for improving image resolution through software technology,which are known as image super-resolution reconstruction algorithms,can be divided into traditional methods and learning-based methods.With the development of computer vision,image super-resolution reconstruction technology based on deep learning is becoming more and more mature.Compared with the traditional algorithm,a variety of high-level features can be extracted with the help of deep learning techniques,thus resulting in better reconstruction quality.However,existing super-resolution reconstruction methods based on deep learning are still suffering from the drawbacks of insufficient receptive fields,high time complexity,and gradient vanishing.To tackle those problems,two convolutional neural network architectures are proposed in this thesis.The main contributions of this thesis are as follows.An image super-resolution reconstruction algorithm based on local and global residual connection convolutional neural network is proposed.Processing images in high resolution space results in the increase of network complexity.Images are therefore processed in low-resolution space to reduce complexity.The reconstruction results are improved by introducing local and global residuals.Local residual promotes the flow of information and avoids the problem of gradient vanishing during network training.Due to the local residual connections,image information can be transferred to deeper layers.The global residual makes the network learn residual information only,resulting in reduced network redundancy and accelerated network convergence.The receptive field is enlarged by increasing the network depth,which makes the network learn more reconstruction information.A super-resolution reconstruction deep learning algorithm based on deconvolution dense block connections is proposed.The convolutional layer and the deconvolution layer are connected in the proposed deconvolution dense block to achieve down-sampling and up-sampling of image features,so that the image information is learned at different stages in the neural network;The extracted features are contacted with deconvolution dense block outputs and there are feature contractions within the deconvolution dense blocks,which prevents the gradient vanishing and increases the network width without introducing any extra parameters.