An Image Feature Classification And Sparse Representation Baesd Super-resolution Reconstruction

In the process of image sampling, image resolution always depends on the condition of thecamera and surroundings, which stimulates the development of image processing. The mostdirectly method to improve image resolution is to increases the pixels or enlarge the size of CCD.But due to the limitations of hardware, researches are mainly focused on software method toimprove the resolution of the images.The interpolation based method, the reconstruction based method and learning basedmethod are the mostly used method for super resolution. Among which, the learning basedmethod is the hot spot in recent years. This method attempts to capture the co-occurrence priorbetween low-resolution and high-resolution image patches. This paper mainly focued onresearching the learning based reconstruction method, which proposed an image-classificationand sparse representation based super-resolution reconstruction method.The main research of this paper are listed below:Firstly, this paper introdued the currently used super-resolution reconstruction technologiesgenerally, and detailed the main idea of each method.This paper highlighted the learning basedSR(super-resolution reconstruction) algorithm.And did researches based on the super-resolutionreconstruction via sparse coding.Secondly, introdueced the principle of feature extraction and classification explicitly,especially focused on texture and gradient extraction.According to the theory of imageclassification, proposed a method to group the images in the training library by different imagefeature before dictionary training. Two dictionaries could be trained from the grouped library.Thirdly, elaborated how to apply the image classification in to the sparse representationbased super-resolution reconstruction.Texture feature extractor and gradient feature extractorwere used to generate two feature image library. Six Gabor filters were applied to extract texturefeatures. Six texture feature training libraries were generated by the Gabor filter with θ of0°，30°，60°，90°，120°and150°. The texture feature dictionaries can be generated by training thesix texture feature libraries. The same to training the gradient feature dictionary, to convolve theimages in training library with eight masks. The gradient feature dictionary could be trained bythe eight feature training libraries. Combined the texture feature libraries and the gradient featurelibraries, and from which could receive the image classification based reconstruction dictionary. Finally, on the research and analysis of super-resolution reconstruction and featureextraction, this paper proposed a feature image and sparse representation based super-resolutionreconstruction.In the process of image reconstruction, divided the image patches into flat partand un-flat part. For the patches in flat part, texture feature dictionary could be used in theprocess of reconstruction. For the patches in the un-flat parts, gradient feature dictionary canused for recovering.This paper focused on applying the image classification into the process ofdictionary training in super-resolution reconstruction. According to the results, dictionaryclassification based super-resolution reconstruction proved to have better reconstructedperformance and improved quality at the edge of the image, besides, the PSNR improved at thesame time.