Reversible Denoising and Lifting Based Color Component Transformation for Lossless Image Compression

An undesirable side effect of reversible color space transformation, which consists of lifting steps (LSs), is that while removing correlation it contaminates transformed components with noise from other components. Noise affects particularly adversely the compression ratios of lossless compression algorithms. To remove correlation without increasing noise, a reversible denoising and lifting step (RDLS) was proposed that integrates denoising filters into LS. Applying RDLS to color space transformation results in a new image component transformation that is perfectly reversible despite involving the inherently irreversible denoising; the first application of such a transformation is presented in this paper. For the JPEG-LS, JPEG 2000, and JPEG XR standard algorithms in lossless mode, the application of RDLS to the RDgDb color space transformation with simple denoising filters is especially effective for images in the native optical resolution of acquisition devices. It results in improving compression ratios of all those images in cases when unmodified color space transformation either improves or worsens ratios compared with the untransformed image. The average improvement is 5.0-6.0% for two out of the three sets of such images, whereas average ratios of images from standard test-sets are improved by up to 2.2%. For the efficient image-adaptive determination of filters for RDLS, a couple of fast entropy-based estimators of compression effects that may be used independently of the actual compression algorithm are investigated and an immediate filter selection method based on the detector precision characteristic model driven by image acquisition parameters is introduced.

💡 Research Summary

The paper addresses a long‑standing problem in lossless image compression: reversible colour‑space transformations based on lifting steps (LS) remove inter‑component correlation but simultaneously spread sensor noise from one component into the others. Because lossless codecs such as JPEG‑LS, JPEG 2000 and JPEG XR rely on entropy coding, any increase in noise directly degrades compression ratios. To solve this, the authors introduce a reversible denoising‑and‑lifting step (RDLS). An RDLS embeds a denoising filter inside a lifting operation while preserving perfect reversibility. The key insight is that the filter is designed to be its own inverse (or to have a symmetric inverse) so that, after the forward transform, the filtered values are passed to the codec, yet during the inverse transform the same filter restores the original noisy values, guaranteeing lossless reconstruction.

The RDLS concept is applied to the well‑known RDgDb colour transformation, which converts RGB into two difference components (R‑G, B‑G) and a luminance‑like component. Simple spatial filters—3×3 mean and bilateral‑median filters—are evaluated as the denoising component of RDLS. Because the filters have a small computational footprint, they can be applied per‑component without increasing the overall complexity of the codec.

A major contribution of the work is the development of fast, entropy‑based estimators that predict the compression effect of a given filter without actually running the full codec. Two estimators are proposed: (1) an entropy‑upper‑bound estimator that approximates the post‑transform bit‑rate from the marginal histograms, and (2) a compression‑effect estimator that measures the change in estimated entropy before and after filtering. These estimators enable an “immediate filter selection” mechanism that chooses the most promising filter for each image on the fly.

Experiments were carried out on three lossless standards (JPEG‑LS, JPEG 2000, JPEG XR) and two image collections. The first collection consists of native‑resolution photographs captured directly from digital cameras and scanners, where sensor noise is most pronounced. The second collection comprises classic benchmark sets such as Kodak and USC‑SIPI. Results show that, for the native‑resolution set, applying RDLS‑RDgDb yields an average compression‑ratio improvement of 5.0 % to 6.0 % across all three codecs. Importantly, the improvement is observed whether the original colour transformation alone would improve or degrade the ratio, indicating that RDLS consistently mitigates noise‑induced penalties. For the benchmark sets, the average gain is more modest but still significant, reaching up to 2.2 % for the best‑performing images.

The paper also introduces a detector‑precision characteristic model that links image‑acquisition parameters (ISO, exposure time, sensor type) to the expected benefit of denoising. By feeding these parameters into the model, the system can predict the optimal filter without any trial‑and‑error, making the approach suitable for real‑time or embedded applications where computational resources are limited.

In summary, the authors demonstrate that:

1. A reversible denoising filter can be seamlessly integrated into lifting‑based colour transforms, preserving perfect reversibility while reducing noise before entropy coding.
2. Fast entropy‑based estimators enable codec‑independent, on‑the‑fly filter selection, eliminating the need for exhaustive compression trials.
3. The RDLS‑enhanced RDgDb transform delivers consistent compression‑ratio gains for both high‑noise native‑resolution images and standard test images across multiple lossless standards.
4. The method scales well computationally and can be driven by acquisition metadata, opening the door to adaptive, sensor‑aware lossless compression pipelines.

Future work suggested includes exploring more sophisticated non‑linear filters, machine‑learning driven filter predictors, extending the technique to other colour‑space transforms such as YCoCg or IPT, and applying the concept to lossless video coding where inter‑frame correlation and sensor noise are also critical concerns. The presented approach promises to improve the efficiency of lossless compression in domains where preserving every bit of image fidelity is essential, such as medical imaging, remote sensing, and archival photography.