Research On Classified Dictionary Learning Super Resolution Reconstruction Algorithm

The high resolution(HR)images contain rich detail information.While,limited by the technologies of the detector processing and the objective factors in imaging,such as camera shaking and system blurring,the images' resolution cannot meet application requirement.Improving the resolution by changing the structure of optical system,or using higher resolution detector,will lead to the increases of system complexity and the processing cost definitely.Improving the resolution by software algorithm is seemed as a good applicability and cost-effective way,which has become the research hotspot in the field of super resolution(SR)imaging.The image SR reconstruction algorithm based on learning calculates the HR images by extracting the features in natural images.Because of the introduction of high frequency information in training samples,the quality of the images reconstructed by the learning based algorithm is significantly improved.However,the traditional SR reconstruction algorithm based on learning calculates the HR images by the same dictionary,which ignores the differences in various images and their different areas,leading to the expansion of reconstruction error.A new SR reconstruction algorithm based on classified dictionary learning is proposed in this article,several pairs of dictionary are obtained by clustering.In the process of reconstruction,the most matching dictionary pair is selected according to the features in low resolution(LR)images,which improves the images' quality without increasing the computational complexity of algorithm.The research work and main innovations include the following:(1)The SR reconstruction algorithm using single dictionary pair is studied and promoted.By establishing the sample sets of natural images,the high frequency component is extracted,which is trained to single HR and LR dictionaries.In the process of image reconstruction,the same dictionary is chosen to reconstruct different LR images.At the same time,two ways to increase the efficiency of the algorithm are adopted in this article: image reconstruction by region management and dimension reduction for sample features.A large number of reconstruction experiments verify the validity of the proposed method.(2)The classified dictionary learning method is proposed.Extracting the multi-dimension characteristics of the pixels in the training samples by Gabor filtering,then the pixels' characteristics are mapped to multidimensional space and the points are clustered by K-means algorithm.The sample images are segmented with the unit of 4?4.Calculating the membership grad of each unit,five groups of sample features are obtained and trained to different dictionaries respectively.(3)The LR images are reconstructed by classified dictionaries.The LR image is divided into patches and then the similarities between the LR patches and each classified dictionaries are calculated separately,the particular dictionary pair with highest similarity is selected for reconstruction.The proposed method improves the reconstruction accuracy without increasing the number of dictionary atoms.At last,the HR patches are joined to a whole image.(4)The SR reconstruction experiments are carried out.To verify the superiority of the proposed SR reconstruction method based on classified dictionary learning,several groups of experiments are designed and the results show that the images reconstructed by the proposed method perform better on both subjective perception and objective measures.Extending the method to the area of infrared image processing,the result is also satisfactory.