Log in

View Full Version : DCHT (Hybrid Discrete Hermite Transform System)


Nania Francesco
7th October 2025, 19:58
My Unexpected Journey

From Spatial Polynomials to a Stable Hybrid Transform (DCHT)

My initial research focused on pushing the limits of lossy compression through spatial domain techniques, specifically exploring adaptive-block polynomial curve-fitting. I painstakingly tested nearly every polynomial formula, aiming for a significant gain in quality, compression ratio, and final visual fidelity—enough to compete with the highly optimized Discrete Cosine Transform (DCT).

However, the results were consistently underwhelming. Despite elaborate adaptive schemes, no polynomial approach could deliver the necessary energy compaction or provide the stability and superior visual performance needed to challenge DCT's dominance. This frustrating plateau made it clear that the future didn't lie in spatial approximations alone.

The Hybrid Breakthrough: Introducing the DCHT
Driven by this roadblock, I pursued a bold new idea: to create a transform that retained the structural efficiency and block-based architecture of the DCT, but relied on a fundamentally different mathematical basis with a crucial spatial-domain linkage to overcome the inherent instability of the DCT when scaled up.

I named this solution the DCHT (Discrete Hermite Transform with a DCT structural call). Its core is a stable hybrid transform, drawing its superior energy compaction and spatial coherence properties from the Hermite-Gauss basis functions. The inclusion of 'DC' in the naming convention (DCHT) is purely a structural nod to the DCT's established two-dimensional compression framework.

The initial results of the DCHT were, frankly, discouraging. The first versions struggled, and the PSNR was initially only half of what DCT achieved, due to poor energy distribution across the AC coefficients.

Through meticulous iteration, focused on rebalancing the weighting of the hybrid components, I found a pivotal point of equilibrium. This balance resulted in a dramatic shift: the DCHT began to concentrate energy far more effectively than DCT.

By using a minimal, fixed number of low-frequency coefficients (196 out of 4096 for a 64×64 block), the new stable transform began to significantly outperform DCT at the same sparsity level.

A Transform in its Own Right
What started as an attempt to fix a problem evolved into a standalone, stable, and reasonably fast transform ready to challenge the DCT era. The DCHT's stability on large blocks, coupled with its superior energy compaction, has established it as a viable candidate for next-generation compression.

Based on these results, I have taken decisive steps to secure its future: the DCHT has been submitted for patent application, its source code has been filed with Zenodo for archival and verification, and most critically, the core mathematical paper has been submitted to the IEEE Journal of Selected Topics in Signal Processing (J-STSP) for peer review."

Columbo
7th October 2025, 22:06
Lotta hype and rhetoric there bro. Can you demonstrate its awesomeness?

Nania Francesco
7th October 2025, 22:39
DCHT: Beyond DCT Limits—Mathematical Proof Attached
Appreciate the demand for proof over promise. We understand the skepticism; the world doesn't need another incremental DCT tweak.

The DCHT (Discrete Hermite Transform) is not a tweak. It's a fundamental change of the transform basis designed to solve the DCT's core weakness: its inefficiency on large blocks (64×64).

The DCT basis forces a circular model onto a discrete rectangular block, wasting entropy. Our DCHT, derived from Hermite-Gauss functions, naturally adheres to the spatial boundaries, achieving dramatically superior sparsity and energy compaction.

The Technical Gravity
To demonstrate this isn't just theory, we present two key facts based on our current experimental program:

Computational Efficiency: Our DCHT prototype, even in its current state, encodes 2.4× faster than our fully optimized SIC reference codec. Speed is not a concern; the DCHT is built for massive parallelization.

Architectural Superiority: The DCHT intrinsically reduces artifacts, eliminating the need for complex, costly in-loop filters (like those used in AVIF) to maintain high visual quality.

The math drives the performance.

Proof of Concept
This theoretical groundwork, including the full mathematical derivation, has been submitted to the IEEE Journal of Selected Topics in Signal Processing (J-STSP) for peer review.

I've attached a PDF excerpt with the fundamental DCHT equations. This is the core basis that makes the difference.

We're not sharing raw benchmark scores yet as the test program is still evolving into a full adaptive codec. But review the math—it speaks for itself.
patent:
https://zenodo.org/records/17288206

Columbo
8th October 2025, 05:12
You can't give a couple screenshots showing the claimed "superior perceptual fidelity"?

Nania Francesco
8th October 2025, 11:27
Of course, I reserve the right to release a closed-source demo of my DCHT-based encoder in the future. For now, I am posting a photo with a comparison between 12% JPEG (I used ImageMagick) and DCHT, which uses 13%. With Ssimulacra2, the comparison is no contest.
JPEG gets -6.89029622
DCHT gets +26.47773956

https://encode.su/attachment.php?attachmentid=12721&d=1759919788
https://encode.su/attachment.php?attachmentid=12722&d=1759937078

Columbo
8th October 2025, 15:28
Thank you. Attachments take forever (or never) to get approved, so please consider an image hosting site instead.

j7n
9th October 2025, 04:28
The quality of these samples is of course below usable, and the differences might not scale to normal compression levels. It seems that overall the resolution of the DCHT sample is reduced. You can see how the logos and signs have become unreadable and there is less detail on the car wheels. Perhaps the JPEG sample could be improved by shrinking it down for a similar resolution loss. The noise gives an impression of detail but it comes in obvious blocks. AVIF clearly wins in this comparison.

Nania Francesco
9th October 2025, 13:19
The quality of these samples is of course below usable, and the differences might not scale to normal compression levels. It seems that overall the resolution of the DCHT sample is reduced. You can see how the logos and signs have become unreadable and there is less detail on the car wheels. Perhaps the JPEG sample could be improved by shrinking it down for a similar resolution loss. The noise gives an impression of detail but it comes in obvious blocks. AVIF clearly wins in this comparison.

Hey, thanks for the honest feedback on those initial DCHT samples. You hit the nail on the head: they looked rough. The resolution reduction and glitches were definitely there, and I appreciate you pointing them out. You're right, AVIF is a clear winner over that buggy initial release.
Those initial samples were generated by a pipeline that had known, annoying bugs: specifically, the Hermite transform implementation was unstable, and the rigid block sizes we were using couldn't properly handle the details.
It was a classic case of quality math (Hermite) stifled by poor design. My fault; I was too eager to show the concept. Bug fixed: the basic Hermite transform is now stable and fully compliant. Smart blocking: I've ditched the rigid structure and implemented finer 16x16 block processing within the larger 64x64 macroblock. This allows DCHT to see and capture those fine details (logos, wheels) without smearing or blocking them.

Columbo
9th October 2025, 16:14
Bug fixed: the basic Hermite transform is now stable and fully compliant. Smart blocking: I've ditched the rigid structure and implemented finer 16x16 block processing within the larger 64x64 macroblock. This allows DCHT to see and capture those fine details (logos, wheels) without smearing or blocking them. Will you post updated samples? Thank you.

Nania Francesco
9th October 2025, 16:59
Jpeg (coded 20%) 213 kb vs DCHT (16x16 blocks) 214 kb
note: DCHT I further compressed it with Jpeg at 88 percent to not take up too much space
https://limewire.com/d/kdkY1#omeJSqPwYu

As you can see I used 16x16 blocks and the deformation effect seems to have disappeared (I had made a mistake in translating my formula into C++)

Z2697
9th October 2025, 21:48
JPEG (1992) uses 8x8 block though, a comparison with 8x8 DCHT will show the real "power" (better or worse) of it.