Research On Compression Image Super Resolution Interpolation Based On Statistical Learning

Image interpolation aims to produce a high-resolution image from its low-resolution counterpart.It is widely used in areas of security monitoring, medical science, HDTV and so on.First of all, this paper describes classical interpolation methods briefly in four categories:polynomial-based, sparse representation-based, autoregressive model-based and direction-based.Then focusing on direction-based interpolation and compression image interpolation, we mainlycomplete the below work:First, analyzing the weakness of modern directional cubic convolution interpolation[1]by meansof theoretical and experimental analysis, we propose a novel directional bicubic interpolation,which decides local edge’s direction and strength depending on gradients in four directions:0°、45°、90°、135°and interpolates with a model of corresponding direction. This method receivesbetter results than algorithm in literature[1].Second, we propose a compression image interpolation method using adaptive symmetricalautoregressive models. This method adopts learning algorithms. So it has two terms: statisticallearning term and image interpolating term. Three innovation points are included in this method:a) Interpolation models. Based on a current autoregressive model, we create a new adaptivesymmetrical autoregressive model which combines the characteristics that traditional interpolationmethods can interpolate some images with good results and images are isotropic.b) Classification in two times. Firstly, classify training images into four directions according tolocal direction and build corresponding training sets. Secondly, classify every direction’s trainingsets into subclasses according to primitive features with K-means clustering algorithm. Thus, theresult of this classification is more accurate.c) Weights of models. The model is chosen whose direction the subclass belongs to. Thenconstrained least square method is used to estimate the weights of the model. At this time, theweights can suit the compression artifacts better and the models can be used to interpolatecompression images better.Third, an adaptive PDE-based regularization preprocessing method about compression artifactsis proposed on the problem that most interpolations can’t deal with compression images well. Theenergy change ratio between primitive and non-primitive areas is used to balance primitivecomponents preservation and compression artifacts removal. Experiments show that methods withthis regularization can interpolate compression images better than methods without regularization.