Efficient Multi-band MLP Based Intra-prediction for Lenslet Image Compression

This paper presents a novel method for lenslet image compression. In the plenoptic representation of lenslet imagery, pixels from adjacent blocks show significantly reduced correlation as compared to what has been observed in natural images. The proposed method first leverages boundary matching to search for the reference blocks of better correlating pixels for intra-prediction. The prediction unit is then decomposed by 2D-DCT into a multi-band representation. It later uses multiple MLP (multi-layer perception) networks for the prediction to explore statistical redundancy and thus reduce the residue for compression.