Adaptive DCT-Domain Down-Sampling and Learning Based Mapping for Low Bit-Rate Image Compression

摘要

This paper proposes a new image compression approach. The high frequency information in an image is adaptively decreased to facilitate compression, which is then compensated through learning-based mapping. Specifically, an adaptive DCT-domain down-sampling method is presented to reduce the high frequencies in images and the visual pattern based learning is introduced to recover the removed high frequencies. The collaboration of the adaptive DCT-domain down-sampling and learning-based mapping presents a new coding mode for image compression (namely D-mode) which can be readily integrated into the mainstream coding schemes. Experimental results demonstrate that our proposed coding scheme achieves better coding performance in terms of both objective and subjective quality in comparison with H.264 intra coding and JPEG 2000, especially at low bit rates.