Block truncation coding using neural network-based vector quantization for image compression

Abstract

A new method is introduced by which a block truncation coder (BTC) is cascaded with a neural network-based vector quantizer (VQ). The proposed coder is very attractive for real time image transmission due to its simplicity and performance. It preserves important characteristics of the image, while cascading the BTC coder with a VQ results in high compression ratios of about 0.5 bpp without significantly increasing the coding time, due to fast coding look-up tables of the VQs. Additional advantages are fast codebook design and reduction of the codebook size required for a given reconstructed image quality.