Sub-block classification using a neural network for adaptive zigzag reordering in JPEG-like image compression scheme
H.-J. Grosse · 1997
A neural network technique for classification of blocks of discrete cosine transform (DCT) coefficients using a backpropagation algorithm is described. The DCT is employed in a variety of transform based image compression schemes. In the authors' recent JPEG like image compression scheme, efficient reordering of coefficients is achieved by applying adaptive zigzag reordering to variable size rectangular sub blocks. The additional neural network based sub block classification discards isolated nonzero coefficients of small significance in some sub blocks and therefore further reduces their sizes. Initial experimental results are presented that demonstrate the potential of the additional neural network based sub block classification in terms of improved coding gain.