Image Super-resolution Reconstruction Based On Sparse Dictionary Learning

The 21st century is the era of information,computer and Internet technology have obtained unprecedentedly rapid development,people have stepped into the era of big data,with the advent of the big data age,a lot of multimedia information appeared on the Internet multimedia,such as high-resolution images,high-definition video,etc.However,the information is inconvenient for storing and transmitting because of the huge needs of memory space and a large quantity.As the need for high-resolution image increases strongly,however,due to the high cost of hardware physical equipment and the limitation of various objective imaging factors,originally obtained image resolution can't reach the desired effect.Therefore the method of improving image resolution through low-cost software has obtained extensive develop space.The image super-resolution reconstruction technique is to reconstruct a high-quality high-resolution image from one or more low-resolution images in the same scene with signal processing technology.Like the speech signal,the image signal is the primary medium for human perception and acquisition of the surrounding environment.According to the latest research,the information obtained through the image signals is about 80 percent of the total amount of information obtained^[1].The traditional Nyquist sampling theorem^[2]requires that the sampling frequency of the signal should be more than twice the maximum frequency of the signal,which could recover the original signal accurately.The limitation of this sampling principle brings more and more difficulties to the collection,transmission and storage of signals.In recent years,Sparse representation^[3]has stimulated the research enthusiasm of many scholars and researchers due to its unique advantages.Sparse representation algorithm is based on the signal sparse representation theory,Any image block on the input low-resolution image can be sparsely represented,by using the per-learned low-resolution dictionary.In this way,the sparse representation coefficient can be calculated,and then,according to the sparse representation coefficient consistency principle,the sparse representation coefficient and high-resolution dictionary are used to calculate the high-resolution image block,which is merged to calculate the high-resolution image block,taking average of overlap between image block,finally we can get the target high-resolution image.Based on this research,the author further improved the algorithm.Adaptive regularization parameters are proposed in the sparse representation phase,which is applied to the stage of dictionary training and image reconstruct.In the feature extraction phase,the features of high and low-resolution images are extracted by using the improved first-order and second-order gradient feature extraction templates,and the high and low-resolution image blocks are obtained.Then,using PCA method^[4],the high dimensional of data obtained is reduced to low the complexity and the time of the dictionary training.In the dictionary learning phase,applying the widely used K-SVD method into the Adaptive regularization parameter to train the high and low resolution image block after the reduced dimension,the corresponding high and low resolution dictionary is obtained.In the image reconstruction phase,through the study of sparse representation model,reconstructing a high-resolution image from a low-resolution image of itself,and then the principle of non-local self-similarity of the image is used to further dehydrate and eliminate artifacts of the reconstructed high-resolution image,so that the reconstructed image can meet people's visual requirement more.