Research On Image Super Resolution Algorithm Based On Convolutional Neural Networks

With the progress and development of society,digital images have been widely used in people's production and life,making people increasingly demanding for the clarity of images.However,due to the production cost,production process,equipment environment and other factors,people often can only get a lower resolution of the image.Therefore,the method of using software to improve image resolution has been widely concerned and applied.This software method is called image superresolution reconstruction.Image super-resolution reconstruction is a method of mapping low-resolution images to high-resolution images by mathematical modeling.Common super-resolution image reconstruction methods can be divided into interpolationbased,reconstruction-based and learning-based three methods.Besides,the method based on deep learning belongs to the method based on learning,and is the current mainstream image super-resolution reconstruction method.After researching and learning the method of image super-resolution reconstruction based on deep learning,we improve the reuse of local features and the strategy of model training,and present a multiconnected convolutional network and two-parameter loss function for image super-resolution reconstruction method.Firstly,based on the idea of densely convolutional network,we design a structure built on multi-connected blocks for the combination of local low-level features and global high-level features.And we build a multi-connected convolutional network using the multi-connected blocks,which enhances the diversity and complexity of feature representation.Then,based on the idea of multi-scale feature fusion,we optimize the architecture of multi-connected convolutional network,and further improve the quality of reconstruction.Finally,a twoparameter loss function is employed to optimize the training process,which makes the reconstructed image contain more abundant details.According to the results of experiments,the proposed method yields state-of-the-art performance both in terms of quality and quantity on five common datasets,indicating an excellent generalization ability and high reconstruction quality.