Research On Image Super Resolution Recontruction Methods Based On Edge Enhancement And Deep Learning

Digital image resolution represents the minimum size of the target that can be resolved by image systems.If the pixel density of an image to the same scene is higher,the resolution of the image will be higher and the contained details will be richer.In the imaging process,due to the limitations of imaging hardware,shooting environment and shooting errors,the resolution of acquired images often fails to meet the actual application requirements.Image super resolution(ISR)is a kind of software-based post-processing methods to enhance image resolution,which can improve the sharpness of images without updating hardwares.It has a wide range of applications such as remote sensing,high-definition television conversion,medical image processing and video monitoring,etc.,and is also one of the research emphases in digital image processing.This dissertation focuses on the interpolation-based and learning-based image super resolution algorithms.In the article,deep researches and analyses are conducted on the theories and methods of super resolution reconstruction.Also,radial basis function interpolation,gradient boosting and convolutional neural network and other related theoretical basis have also been introduced.The aim of this dissertation is to apply the radial basis function and learning algorithms with strong generalization ability to image super resolution reconstruction.And in this way,the edge and texture details of the upsampled images are preserved as well as high frequency information is introduced,which results in generating visually appealing super resolved images.In this dissertation,three super resolution algorithms are proposed for different application requirements:(1)local adaptive based image interpolation method;(2)gradient boosting based image super resolution method;(3)depthwise separable convolution based image super resolution method.In the meantime,natural images and magnetic resonance(MR)images are used to verify the super resolution reconstruction ability of the algorithms.The main contents of this paper are as follows:1)In this dissertation,a locally adaptive image interpolation method is proposed.This method uses LPA-ICI algorithm to search for the optimal shape of the neighborhood of the HR pixels to be interpolated,and then the radial basis function is utilized to fit their pixel values.Further more,guided image filter is employed to enhance the edges and texture details of the interpolated images,so the visual effect of images is improved.The advatages of the algorithm is simple and the model doesn't need training.This method is suitable for super resolving images without training data.2)Aiming at the problems of weak generalization ability and higher reconstruction errors of traditional learning-based super resolution algorithms,this dissertation introduces gradient boosting to deal with super resolution reconstruction.Firstly,the gradient boosting framework is extended into multiple outputs mode,and decision tree is taken as weak learners.The decision trees are trained in a sequential mode,which can reduce the bias and variance of the model.Secondly,shrinkage strategy is utilized to enhance the generalization ability of the model.Meanwhile,subsampling is employed to reduce the number of training samples and decorrelate the decision trees as well as accelerate the training speed.Finally,in testing,the independent prediction results of each tree are linearly weighted and combined to produce high-resolution images.Experimental results show that the proposed algorithm is more accurate and the details of the super-resolved images are sharper than traditional learning-based algorithms.This method can achieve higher super resolution performance with a smaller training set,and the training speed is relatively fast.So it is suitable for image super resolution with a relative small training set.3)Aiming at the problems of high complexity,difficulty in training and slow convergence of deep super resolution algorithms,this dissertation proposes a depthwise separable convolutional network based image super resolution model.Firstly,depthwise separable convolution layer is employed to replace the conventional convolutional layer in the proposed model and the normalization layer is removed,in this way the number of parameters is reduced.Moreover,the features are extracted directly from low-resolution images,and then transposed convolution is employed to map low-resolution features into high-resolution feature space.So the complexity of the model is further decreased.Secondly,densely skip connections and residual network structure are employed to improve its robustness.Finally,extensive experiments have been designed to choose the hyper-parameters,including the number of modules and convolutional layers in one module.Experimental results show that this algorithm has strong generalization ability,and the objective evaluation indexes and visual effects of reconstructed images are better than other algorithms.This method is suitable for image super resolution applications with large-scale training set and higher accuracy requirements.