| High resolution images can not only bring us a better intuitive visual experience,but also help to improve the accuracy of data analysis in many application scenarios.Therefore,high resolution images are widely used in medical imaging,security monitoring,satellite remote sensing,military reconnaissance,high-definition television and other fields.However,in the process of digital image acquisition,the scene information will be affected by a series of degradation factors,so it is difficult to obtain high-resolution images directly,which can not meet the actual application requirements.Increasing the sampling density of the sensor array of the acquisition equipment or changing the structure of the optical system to improve the imaging resolution will lead to the multiplication of equipment cost and system complexity,which limits the scope of its practical application.Image super-resolution(SR)technology can use algorithms to improve the resolution,bypass the limitations of imaging system hardware,with higher cost performance and a wider range of applications,so it has been a research hotspot in the field of image processing at home and abroad.The image SR algorithm based on learning can use the features of natural images in the sample database as a priori information to supplement high-frequency details for low resolution images,which can effectively improve the image resolution.Therefore,this thesis is based on this kind of algorithm and there are some main stages of research as follows.Firstly,research the learning based super-resolution algorithm and its ideas and methods improved.This thesis studies the principle of SR algorithm,and improves the algorithm based on sparse representation theory and structural similarity constraints that commonly used in SR field.Meanwhile,this thesis implements a super-resolution algorithm based on regional analysis,which uses the function model to directly learn the corresponding SR mapping relationship after the integration of features,and then construct the image by weighting the SR mapping model according to the results of regional analysis.Secondly,design and implement the regional feature integration method based on fuzzytheory.In order to ensure that the mapping relationship between high-resolution and low-resolution images can be accurately learned in the training stage,this thesis introduces the fuzzy theory to quantitatively analyze the image regions in the form of membership degree and integrate the regional feature information.This processing method can ensure the consistency of the feature information used in the training and improve the data prediction accuracy in the reconstruction process.Thirdly,design the SR mapping relationship learning method based on mixture of Gaussian process model.In order to save the calculation cost of SR reconstruction,this thesis uses the function model to learn the SR mapping relationship directly.At the same time,in order to improve the accuracy of the studied mapping relationship,this thesis designs a mixture of Gaussian process model to train the model parameters,aiming for reducing the deviation of the single model learning results.Finally,carry out the experimental verification and the result analysis of super-resolution reconstruction.In order to verify the effectiveness of the SR algorithm based on regional analysis designed in this thesis,a lot of standard images and real images are used for image reconstruction experiments,and the results are analyzed and compared.At the same time,during the process of real image testing,the thesis extends the algorithm based on the degradation model.The experimental results show that the algorithm in this thesis can not only achieve better reconstruction effects but also effectively improve the reconstruction speed. |
