Research On Fast Super-resolution Technology Based On Local Regression

The purpose of super-resolution (SR) is to recover a high resolution image from oneor more low resolution images. Because of its low cost and ability to improve the imagingresults of the widespread low-quality imaging sensors，SR Has been extensively studied inrecent years. In this paper, we focus on the example-based single-frame image SR problem,proposed a fast super-resolution algorithm based on local regression, we also provide afurther accelerated method of this algorithm. The quantitative analysis of thesuper-resolution results show t the effectiveness and efficiency of our method.Example-based SR has received great research interesting since it has been proposed,this method use an external database to predict the missing high-frequency information inlow resolution images. We motivated by the idea of local learning and assume themapping from mid-frequency space to high-frequency space is smiliar locally.Mid-frequency infomation vector in the database has been divided into different categoriesin the trainging phase, without the searching implemt in the SR phase, we cansignificantly improve the speed of our super-resolution algorithm. In addition, weanalysed the relationship between the Mid-frequency infomation energy and the missinghigh-frequency information, and only perform super-resolution operation on theMid-frequency information vector that contains a large energy to further enhance thespeed of the algorithm.In the experimental section, we first verify the correctness of our hypothesis aboutthe local similarity of mapping functions, and analysed the effect of the database size andcategories numbers on our method; then the acceleration performance of our speed-upversion algorithm are tested on different images; Finally, we compared our algorithm withother algorithms, the results of the comparison show that our method can generate highquality super-resolution results efficiently.