High-performance discrete wavelet transform for JPEG XS standard

Abstract

This paper presents a high-speed method for forward and inverse discrete wavelet transform (DWT) intended for the JPEG XS image compression standard. Unlike state-of-the-art approaches, which process pixels sequentially, the proposed algorithm employs the Winograd method to compute groups of 2-5 pixels in parallel within a single clock cycle. We determine the minimum fractional bit-widths required for fixed-point arithmetic to ensure reconstructed image quality with a peak signal-to-noise ratio (PSNR) of at least 40 dB. Hardware modeling using the OpenLane environment demonstrates that the proposed method increases throughput by up to 109% for forward DWT and up to 144% for inverse DWT compared to state-of-the-art techniques. The optimal configurations are 3-pixel fragments for forward and 4-pixel fragments for inverse transforms. The proposed DWT approach is recommended for real-time systems where processing speed is critical, particularly in medical imaging and satellite data processing.