Single Image Super-Resolution Reconstruction Based On Random Forests

In the modern information society,image plays an important role as the carrier of transmitting information,and images with higher resolution can provide rich details and thus have higher application values.Due to the technological limitations of digital imaging systems,the images acquired are usually of lower resolutions.In this context,the concept of image super-resolution reconstruction arises at the historic moment,and its goal is to reconstruct clear high-resolution images from the blurred and degraded low resolution ones.At present,super-resolution reconstruction algorithms have been widely used in security,medical,video and remote sensing fields.In this paper,we firstly discuss the related theory and technical basis of image super-resolution reconstruction,and the research results in this field both at home and abroad,and focuses on learning based super-resolution reconstruction algorithm.The main work of this paper contains the following 4 aspects:(1)Image super-resolution reconstruction algorithm based on random forests: during the training phase,external training data are used to train a random forests containing multiple decision trees,the leaf nodes of the decision tree for learning linear mapping relationship between the low resolution image patches and the corresponding high resolution image patches,namely regression model;during the reconstruction phase,the regression models of the leaf node corresponding to the input image patches are searched to predict the high resolution image patches.Finally,the prediction results of the multiple trees are combined to reconstruct the high resolution image.(2)Increase training data to improve the quality of reconstructed images: we extend our training data by appropriately increasing the open-source image data sets,and compare the quality of the reconstructed images before and after data increment.(3)Improve the quality of reconstructed images by enhanced prediction method: we get 8 pieces of low resolution images through different rotation angles and the corresponding flipping,apply super-resolution reconstruction to each image,then reverse the transformation,and average 8 image reconstruction of high resolution image as the last.(4)Improve the quality of reconstructed image by hierarchical learning: the decision tree is no longer independent of each other,but is a hierarchical relationship.The high-resolution image patches predicted by each layer of decision tree go into the next layer of decision trees to continue prediction,and the hierarchical push prediction of high resolution image patches are close to the real high resolution image patches.In image reconstruction stage,the decision tree in each layer drives the predicted high resolution image patches to approach the real high resolution image patches gradually.Lots of experimental results show that the proposed method has very good performance in image reconstruction quality and reconstruction efficiency.