Image-based rendering VIA learning of the block dictionary

In this paper, we proposed an image-based rendering method based on the block sparsity of epipolar plane image(EPI). This method considers the structural similarity and block sparseness features of scene signal described by EPI model, which makes it possible to estimate approximately original signal with less measured values, and reduces the complexity of signal sampling and processing. First, we choose a part of scene EPI images as samples and carry on sub-block processing. Furthermore, we train the sparse representation dictionary by KSVD method, and use the low-frequency part of Discrete Cosine Transformation (DCT) matrix to build the sampling matrix. Finally, we reconstruct original signal by orthogonal matching pursuit method. According to the results of simulation experiment, compared with traditional methods, under the same sample rate, the proposed method promotes both the subjective and objective quality of complex scene with regard to different depths.