Image Reconstruction Using Semi-supervised Learning On Big Data

Impulse noise corruption in digital images frequently occurs because of errors generated by noisy sensors or communication channels,such as faulty memory locations in devices,malfiunctioning pixels within a camera,or bit errors in transmission.Although recently developed big-data streaming enhances the viability of video communication,visual distortions in images caused by impulse noise corruption can negatively affect video communication applications.This paper presents a novel model that uses a devised cost function involving semi-supervised learning based on a vast amount of corrupted image data with a few labeled training samples to effectively remove the visual effects of impulse noise from the corrupted images.The main work includes the followings:(1)The establishment of Highly-accurate image reconstruction model.When an image is corrupted by high-density noise,clean pixels are limited since most of the pixels in the image are covered by noise.This paper proposes to first extract the global context,in which all the clean pixels in the image are used to predict the entire image,and then extract local context and social context to further restore the image details.The model considers the global context,local context and social context synthetically,by which the images corrupted by high-density impulse noise can be effectively and accurately reconstructed.(2)Cost function designing for model optimization based on the existing issues under big image data.As the increasing viability of more video communication services,there are three problems existed across large volumes of noisy images:high-density noise,sparse samples,and the noise multimodality.This paper proposes the cost function in three stages according to these three problems.In the first stage,the cost function term for high-density corrupted image restoration is proposed by employing image spatial correlations.In the second stage,the cost function term based on semi-supervised learning is proposed for the sparse training samples.In the third stage,we introduce the cost function term for the noise multimodality by using properties of natural images.(3)Model training based on a semi-supervised learning technique.After the model is proposed,the focused study of this paper is transformed into an ill-posed issue,where an optimal solution satisfying the cost function is attempted to be found out.In the case where the model is known,the parameters are unknown,and the cost function is complex,maximum likelihood estimation method is considered to effectively estimate the parameters of the model.A likelihood function derives from the cost function designed in this paper is used to construct the learning strategy,so that the optimal model can be efficiently obtained.In the experiments,this paper conducted the image reconstruction comparison on both single highly-corrupted image and large image sets corrupted by multiple impulse noise.The results show that proposed model outperformed the existing state-of-the-art image reconstruction models in terms of both image gray values and image structures.