[Paper Review] New Transforms for JPEG Format

Stanislav Svoboda, David Barina|arXiv (Cornell University)|May 9, 2017

Advanced Data Compression Techniques13 references3 citations

TL;DR

This paper evaluates alternative transforms—such as the local cosine transform (LCT), discrete wavelet transform (DWT) with CDF 9/7 wavelets, discrete Chebyshev transform (DChT), and Walsh–Hadamard transform (WHT)—as replacements for the standard 2D DCT in the JPEG compression pipeline. The LCT outperforms DCT at low bitrates by reducing blocking artifacts, while the CDF 9/7 DWT matches DCT performance at high bitrates.

ABSTRACT

The two-dimensional discrete cosine transform (DCT) can be found in the heart of many image compression algorithms. Specifically, the JPEG format uses a lossy form of compression based on that transform. Since the standardization of the JPEG, many other transforms become practical in lossy data compression. This article aims to analyze the use of these transforms as the DCT replacement in the JPEG compression chain. Each transform is examined for different image datasets and subsequently compared to other transforms using the peak signal-to-noise ratio (PSNR). Our experiments show that an overlapping variation of the DCT, the local cosine transform (LCT), overcame the original block-wise transform at low bitrates. At high bitrates, the discrete wavelet transform employing the Cohen-Daubechies-Feauveau 9/7 wavelet offers about the same compression performance as the DCT.

Motivation & Objective

• To investigate whether alternative transforms can outperform the standard DCT in the JPEG compression pipeline.
• To assess the impact of transform choice on image quality, particularly in terms of blocking artifacts and PSNR.
• To compare the performance of block-wise transforms (like DCT) with overlapping or wavelet-based alternatives.
• To identify the most effective transform for maintaining image quality across varying compression rates.
• To explore the feasibility of replacing DCT in JPEG without modifying the rest of the compression chain.

Proposed method

• The study uses a standard JPEG compression pipeline with DCT replaced by alternative transforms, preserving quantization and entropy coding stages.
• Each transform is applied to 8×8 blocks of YCbCr components after chroma subsampling and DC-level shifting.
• The peak signal-to-noise ratio (PSNR) is used as the primary metric to compare image quality across transforms.
• The local cosine transform (LCT) is implemented as an overlapping, block-adjacent variant of DCT to reduce blocking artifacts.
• The discrete wavelet transform (DWT) employs the CDF 9/7 and CDF 5/3 biorthogonal wavelets, with separable and non-separable implementations tested.
• The discrete Chebyshev transform (DChT) and Walsh–Hadamard transform (WHT) are evaluated using their standard definitions and implementations.

Experimental results

Research questions

• RQ1 Can the local cosine transform (LCT) improve image quality over the standard DCT in JPEG at low bitrates?
• RQ2 Does the discrete wavelet transform with CDF 9/7 wavelets match or exceed DCT performance at high bitrates?
• RQ3 How do non-block-based transforms like LCT and DWT compare to block-based DCT in reducing blocking artifacts?
• RQ4 Does the discrete Chebyshev transform (DChT) or Walsh–Hadamard transform (WHT) offer any advantage over DCT in PSNR or visual quality?
• RQ5 Can the JPEG compression pipeline be effectively adapted to use transforms other than DCT without re-architecting the entire system?

Key findings

• The local cosine transform (LCT) outperforms the standard DCT at low bitrates, achieving a PSNR improvement of 0.3 dB in visual comparisons.
• The LCT effectively reduces blocking artifacts, as demonstrated in visual comparisons between DCT and LCT outputs.
• At high bitrates, the separable CDF 9/7 discrete wavelet transform (DWT) delivers compression performance comparable to the DCT.
• The non-separable CDF 9/7 DWT produces visible stain artifacts on object edges, making it inferior to the separable variant.
• The discrete Chebyshev transform (DChT) consistently underperforms the DCT across all tested conditions, with PSNR values slightly below DCT.
• The Walsh–Hadamard transform (WHT) offers no PSNR advantage over DCT and is only beneficial for computational speed.