Wavelet Packet Based Video Super-resolution Using Radial Basis Function Neural Network

Video super-resolution uses a set of successive low-resolution (LR) images in a video sequence to reconstruct a high-quality and high-resolution (HR) image. In this paper, a video super-resolution image reconstruction method using local full search (LFS) motion estimation algorithm, wavelet packet transform (WPT) and radial basis function (RBF) neural network is proposed. LFS motion estimation algorithm is used to collect motion-trace volume which can eliminate deep object motion in the LR images. The HR input image is decomposed by using two-level two-dimension (2D) WPT. The resulted sub-images are employed to train the networks. RBF neural network is applied to each sub-band to learn the relationship between LR and HR images. The super-resolved (SR) image is finally produced by using the inverse WPT. Experiments are mainly focused on image sequences from real videos with different kind of motions. The objective and subjective quality assessments are carried out and compared with the conventional super-resolution methods. The experimental results give that the proposed method outperforms the conventional and recent super-resolution methods for real video sequences. It is well noticed that the proposed method has the best results in terms of both the objective quality and visual quality of the super-resolved image.