Research On Image Super-resolution Based On Broad Learning System

With the development of information technology,digital images are widely used in various fields of production and life.High-resolution images are rich in color and detail information,which not only provides a good visual experience but also facilitates various subsequent image processing tasks.Image super-resolution technology aims at using software technology to reconstruct high-resolution images from multiple or single low-resolution images.In the past few decades,image super-resolution technology has been applied in many fields,such as surveillance,medical diagnosis,bio-information recognition and astronomical observation.The broad learning system is a new tool proposed in the field of machine learning in recent years,which is mainly based on a flat network model with a simple structure and few parameters.At present,the broad learning system has produced good results in tasks such as image classification and face recognition.This paper studies the image super-resolution technology based on the broad learning system.The main contents are as follows:First,an image super-resolution algorithm based on broad learning is proposed.During the training phase,the algorithm takes low-resolution blocks as input to the broad neural network.At the same time,in the feature layer of the broad learning system,the input low-resolution image is further enhanced.Therefore,the algorithm actually establishes a non-linear relationship between low-resolution image blocks as well as their potential features and high-resolution image blocks,making full use of the information contained in low-resolution images,and improving quality of reconstructed image.In the reconstruction stage,high-resolution images can be obtained by directly inputting low-resolution images to the trained network.This algorithm uses a least squares algorithm to train the network and does not require iterative soluting process.It avoids falling into local optimization and reduces the complexity of the algorithm,which significantly reduces the computing resources.The experiments show that the algorithm can achieve better reconstruction results while ensuring faster training time.Secondly,in order to effectively recover high-frequency detail information in images,an image super-resolution algorithm based on global residual learning and broad learning is proposed.This method takes low-resolution image blocks as input,and uses a broad learning network to establish a mapping between low-resolution image blocks and high-resolution residuals.Since this method only uses the high-frequency part of the image as training sample,the amount of data used is reduced,the training speed is faster,and a reconstructed image with richer details is obtained.Experiments show that this method not only improves the training speed of the model,but also improves the high-frequency parts such as texture and edges of the reconstructed image.Finally,in order to enhance the robustness of the broad learning network to noisy and outlier training data,a broad learning network based on the correntropy criterion is proposed.The network replaces the traditional mean square error loss function with correntropy,and a corresponding learning algorithm is given.The improved broad learning network is used for image super-resolution reconstruction,and the mapping between the low-resolution image and the high-resolution image is learned.Experimental results show that the algorithm significantly improves the robustness of the broad learning network and improves the quality of reconstructed image.