The goal of the super-resolution algorithmis to convert the input low-resolution image to a high-resolution image. Highresolution means that the pixel density in the image is high, andhigh-resolution image can provide more details which is indispensable in manypractical applications, and these details are of importance in manyapplications, such as surveillance, medical imaging, satellite imagery, video,or other specific area. Due to the low – resolution to high – resolutionsolutions are not unique, it is a typical ill – posed problem in imageprocessing. To solve this typeof problem, we need some powerful priori knowledge to make a robust estimate.
For the evaluation of super-resolution algorithm, the degree of similaritybetween the predicted result and the actual situation in the experiment is animportant index to measure the super-resolution.At present, the most commonly used approach to super-resolution isinterpolation. Interpolation is a method of redistributing pixels of anoriginal image to change the number of pixels. Such algorithm has a smallamount of calculation and simple structure, so the running speed is very fast.However, the grayscale value after the resampling has obvious discontinuities,and the image quality is greatly lost, resulting in obvious mosaic and jaggies.Therefore, the interpolation algorithm is insufficient to meet the needs ofsome areas, so we will introduce some more complicated and effective methods inthe follow-up.An example-based method is currently most popular super-resolutionalgorithm that through the image patches, to establishment of prediction modelsto get the target image.
These image patches may come from an external databaseimages or input image, the former needs to extract many image patches from the database.The latter through the image self-similarity, i.e., local image structures tendto recur within at different image scales.
In this paper, we propose a method for super-resolution on the fronthuman face in a single channel image. By using the similarity of front faceimages, we can establish a locally weighted regression that can predict thehigh-frequency image patches from the database and input images, and thetarget high-resolution image is obtained by combining the input low-resolutionimage patch and the predicted high-frequency image patch. Our proposedalgorithm (SRLWR) has several attractive advantages: (1) Compared with othersuper-resolution algorithms, our algorithm has a high degree of restoration,even in the case of more blurred input images. (2) Simple structure, the use of facial features of the high similarityof images, we only need to establish high-resolution and low-resolution patchcorresponding locally weighted regression. Fig. 1 shows the framework of SRLWRalgorithm. (3) A more intuitiveimage evaluation method has been designed to facilitate the analysis andcomparison of the predicted images. (4) Adding eye alignment to pre-processingcan reduce output image artifacts.
The remainder of this paper is organized as follows. Section II isour related work. Section III will describe in detail the implementation ofSRLWR algorithm. Section IV is the experimental part, through the experimentselect the appropriate experimental parameters, and compared with the existingadvanced algorithms. Section V concludes this paper. Notation: H, X and Y denotes thehigh-frequency image, low and high-resolution image respectively, in themiddle, represents the result of enlarging X. h, x and y denotes the corresponding imagepatch respectively, and for training and predicted part we use subscripts torepresent it, like , .
For the definition of the parameters, P and Orepresents the patch size and overlap size, and use F to represents ourproposed SRLWR algorithm.