Image Super-resolution Via An Improved Sparse Representation Method

The objective of Image Super-Resolution (SR) is to produce a high-resolutionimage from the low-resolution image which captured from the same scene. Thistechnique makes it easier to distinguish the interested target objects or persons viazoom out the interested region of target image.This technique is widely used in manyvisual applications, such as surveillance, satellite image analysis, medical diagnostics,and high-definition TV etc. Although sparse representation based image SR is one ofthe best approaches in this field, it still insufficient in image quality and processingspeed.This paper mainly focuses on improving the quality of the reconstructed imageand the processing speed, based on the image SR via sparse representation method.Firstly, Theoretical analysis on the factors which effect the efficiency of thefeature-sign algorithm and the experimental results show that we can speed-up thealgorithm by reduced the dictionary size or given the initial value of the sparsecoefficient. Then, this paper proposesa SR method based on feature classification. Asthe single feature cannot reconstruct the image contour efficiently, one supervisedfeature extraction approach was proposed to extract the edge and texture features fromthe first-order and second-order gradient images marked by the binary edge imagebycanny algorithm. Using the dual features, we improve the image SR quality and thespeed. To speed-up the algorithm, we design a feature cluster algorithm before thesparse coding stage.This method will obtain more structural and lower dimensionalsub-dictionaries. Therefore, we can select the most relevant sub-dictionary for eachinput low-resolution image patch to reconstruct the high-resolution image patch, so as toimproving the processing speed and the image quality. Finally, the overlap-basedcluster method was proposed to balance the independence and compatibility betweeneach sub-dictionary, this cansolve the reconstructed information scattered problemleaded by the cluster. The experiments demonstrate that the proposed method in thispaperimproves not only the quality of the reconstructed image, but also the processingspeed.