Video super-resolution by integrating SAD and NCC matching criterion for multiple moving objects

Traditionally, image enlargement is magnified from a single image. Due to the one and only one image, the quality of the reconstructed image is thus constrained. Super-resolution is proposed to use multiple frames as additional information to estimate the high-resolution image. If we have enough low-resolution images with observed subpixels, the high-resolution image can be reconstructed. In this paper, a complete super-resolution model based on k-means motion clustering is proposed for image enlargement with multiple moving objects. Motion masks are first produced for useful image selection and then blocks matching are used to do motion estimation. Two complementary features, sum of absolute difference (SAD) and normalized cross-correlation (NCC), are adopted as the matching criterion. Objects are assumed to move slightly between two consecutive images. Thus, erroneous motion vectors could be corrected by the center of motion clusters. The proposed method achieves better magnification quality than the traditional ones and some previous super resolution works. From the experimental results, both the visual and quantitative improvements, especially in the high frequency, are significant.