Variable dimension quantization in the transform domain

Quantization is a critical procedure in lossy image compression. Linear (uniform) quantizers are used in most of the present image (sequence) compression standards for their simplicity and flexibility. Unfortunately, the resulting representation is often inefficient, preventing potentially significant gains in image compression. We introduce a new class of linear, low complexity, variable dimension quantizers (VDQ) that efficiently reduces the redundancy of the data stream while retaining the design flexibility and signal-to-noise performance of a linear quantizer. When applied to frequency domain representations in the JPEG image compression standard, this approach resulted in SNR improvements of almost 9 dB over the baseline quantizer at comparable bit rates. The algorithm presented is near optimal and runs in O(N) time, making it suitable for real time applications. Its low complexity and effectiveness make VDQ a promising alternative to conventional quantization for image codecs