Learning-based super resolution using kernel partial least squares

a r t i c l e i n f o In this paper, we propose a learning-based super resolution approach consisting of two steps. The first step uses the kernel partial least squares (KPLS) method to implement the regression between the low-resolution (LR) and high-resolution (HR) images in the training set. With the built KPLS regression model, a primitive super-resolved image can be obtained. However, this primitive HR image loses some detailed information and does not guarantee the compatibility with the LR one. Therefore, the second step compensates the primitive HR image with a residual HR image, which is the subtraction of the original and primitive HR images. Similarly, the residual LR image is obtained from the down-sampled version of the primitive HR and original LR image. The relation of the residual LR and HR images is again modeled with KPLS. Integration of the primitive and the residual HR image will achieve the final super-resolved image. The experiments with face, vehicle plate, and natural scene images demonstrate the effectiveness of the proposed approach in terms of visual quality and selected image quality metrics. In the applications of surveillance, object tracking, and vehicle license plate recognition, the acquired images are usually of low resolution, which may result in a failure of further analyses such as segmentation and recognition. Deriving a high-resolution (HR) image from the low-resolution (LR) one or a sequence of LR images provides a solution to these applications, which is known as the super resolution (SR) imaging technique [3,17,25]. The derived high-resolution image is also called super-resolved image. Generally, the SR algorithms can be categorized into two classes, i.e. multi-frame based approach [7,10,18] and single-frame based approach, which is also called learning-based approach [1,5,13,14,29]. In the multi-frame based approach, the HR image is derived from several LR observations of the scene, which are typically aligned with sub-pixel accuracy; while in the learning-based approach an image database, which includes LR and HR image pairs, is used to infer the HR image from its corresponding LR input. The basic idea of learning-based approach is to model the relation between LR and HR images with the available image pairs in the database and then infer HR image from input LR image with the established model. Compared with the multi-frame based approach, which basically processes images at the signal level, the learning-based approach incorporates more prior information to infer the unknown HR …