How2Compress: Quality-Constrained Image Compression

• How2Compress is an automated tool for guaranteed-quality lossy compression of grayscale optical images using iterative parameter control.
• It supports multiple coders and full-reference quality metrics such as PSNR-HVS-M and MSSIM to ensure visual fidelity under strict quality constraints.
• By integrating component-wise processing and dynamic-range normalization, the tool enables optimal coder comparison and parameter tuning for diverse imaging applications.

How2Compress is an automated software tool for guaranteed-quality lossy compression of grayscale optical images. It is designed to compress single-channel images by a selected coder under a selected full-reference quality metric, while enforcing a preset target value within a user-defined tolerance. The system supports optional component-wise processing of multichannel or hyperspectral data, auxiliary format conversion such as RAW, and optional direct or inverse homomorphic transformations for dynamic-range normalization of hyperspectral sub-bands. Its core purpose is to obtain the maximal compression compatible with a prescribed quality constraint and to enable direct comparison of coders under identical metric targets (Krivenko et al., 2024).

1. Scope, data model, and operating objective

How2Compress operates primarily on grayscale optical images. Its stated input is a single-channel image, although the workflow can be extended to multichannel or hyperspectral data by compressing each component or sub-band separately. In hyperspectral settings, preprocessing can include format conversion and dynamic-range normalization by homomorphic transforms when different bands exhibit very different value ranges (Krivenko et al., 2024).

The tool is organized around a constrained quality-control problem rather than around a fixed bitrate target. For a coder parameter $\theta$ and bitrate $R(\theta)$, the operational objective is

$$\min_\theta R(\theta) \quad \text{subject to} \quad Q(I,\hat I(\theta)) \ge q_0 \quad \text{and} \quad \theta \in [\theta_{\min},\theta_{\max}],$$

where $Q$ is the chosen quality metric, $q_0$ is the target quality, and $\hat I(\theta)$ is the reconstructed image (Krivenko et al., 2024).

This formulation places How2Compress within the rate–distortion tradition, but with the constraint expressed directly in terms of a user-selected fidelity or HVS-oriented metric. A plausible implication is that the tool is intended less as a new codec than as an orchestration layer for coder comparison, parameter search, and reproducible guaranteed-quality operation.

2. Software architecture and control workflow

The graphical interface is described as having eight functional blocks: Coders’ list, Metrics’ list, Quality, Quality accuracy, QS_min/QS_max or bpp bounds, control buttons, input/output image display, and quantitative results (Krivenko et al., 2024). The overall workflow proceeds by selecting an image, optionally applying preprocessing, choosing a coder and metric, specifying a target quality, defining parameter bounds, estimating the feasible metric range, and then launching an automated control loop.

For coders controlled by quantization, the user specifies bounds as QS_min and QS_max. For SPIHT and JPEG2000, the controlled variable is not a quantization step but output bitrate in bits per pixel, with Min/Max bpp in the range 0 to 8. The “Min/Max Estimate” function compresses at the parameter extremes and computes the corresponding metric values. This establishes feasibility and initializes the subsequent search (Krivenko et al., 2024).

The decision logic is explicit. The user selects a coder, and the tool does not auto-switch coders mid-run. Instead, for the chosen coder, the system automatically adjusts its main control parameter: quantization scaling factor for DCT-like coders, or bpp for SPIHT and JPEG2000. The output view presents the input and reconstructed images side by side and reports the attained quality, the number of iterations, the final control parameter, and the compression ratio. For hyperspectral data, “Start hyperspectral” triggers sequential sub-band processing with per-band reporting (Krivenko et al., 2024).

This workflow is centered on monotone parameter control. The tool explicitly relies on the fact that increasing the quantization scaling factor coarsens quantization, raises compression ratio, and degrades the quality metrics, while decreasing bpp in SPIHT or JPEG2000 likewise increases compression and reduces quality.

3. Supported coders and quality metrics

How2Compress supports seven coders: JPEG, SPIHT, JPEG2000, AGU, ADCT, AGU-M, and ADCT-M. JPEG is described as a baseline block-DCT coder; JPEG2000 and SPIHT are wavelet-based coders parameterized by output bpp; AGU and ADCT are high-quality DCT-based coders; and AGU-M and ADCT-M are HVS-oriented modifications of AGU and ADCT that are “able to take into account some peculiarities of the human visual system” (Krivenko et al., 2024).

The metric set comprises PSNR, PSNR-HVS, PSNR-HVS-M, MSSIM, and WSNR. For PSNR, PSNR-HVS, PSNR-HVS-M, and WSNR, targets are specified in dB. For MSSIM, the target lies in $[0,1]$. The tool enforces the constraint $Q(I,\hat I) \ge q_0$ for these increasing metrics, with an accuracy tolerance $A$ such that the achieved metric falls within $q_0 \pm A$ (Krivenko et al., 2024).

The definitions given in the description are standard. Mean squared error is

$R(\theta)$0

PSNR is

$R(\theta)$1

SSIM for patches $R(\theta)$2 and $R(\theta)$3 is given as

$R(\theta)$4

and the implementation uses MSSIM through the Metrix MUX package with default constants $R(\theta)$5. PSNR-HVS and PSNR-HVS-M are full-reference HVS-based metrics in dB, with PSNR-HVS-M additionally accounting for contrast masking, while WSNR is an HVS-weighted SNR also reported in dB (Krivenko et al., 2024).

For practical selection, the paper recommends PSNR-HVS-M or MSSIM when visual quality matters. It further states that PSNR-HVS-M around 40 dB or MSSIM around 0.985 corresponds to a “distortion invisibility threshold,” described as practically visually lossless (Krivenko et al., 2024).

4. Iterative quality control and parameter search

Two automated strategies are implemented. The first is median split, effectively a binary search over the control-parameter interval. The second is a recommended-start strategy with linear interpolation at the final stage (Krivenko et al., 2024).

In median split, the tool initializes an interval $R(\theta)$6, repeatedly evaluates the midpoint, compresses and decompresses at that parameter, computes the metric, and shrinks the interval according to whether the result is above or below the target. For monotone decreasing $R(\theta)$7 with increasing $R(\theta)$8, if $R(\theta)$9, the tool moves toward higher compression; if $$\min_\theta R(\theta) \quad \text{subject to} \quad Q(I,\hat I(\theta)) \ge q_0 \quad \text{and} \quad \theta \in [\theta_{\min},\theta_{\max}],$$0, it moves toward lower compression. The process stops when $$\min_\theta R(\theta) \quad \text{subject to} \quad Q(I,\hat I(\theta)) \ge q_0 \quad \text{and} \quad \theta \in [\theta_{\min},\theta_{\max}],$$1 or the interval becomes sufficiently small (Krivenko et al., 2024).

In the recommended-start method, the search begins from a heuristic parameter estimate that depends on coder and metric. The tool then performs a small number of compression–metric evaluations to bracket the target and computes the final parameter by linear interpolation between the nearest trials that straddle the desired metric:

$$\min_\theta R(\theta) \quad \text{subject to} \quad Q(I,\hat I(\theta)) \ge q_0 \quad \text{and} \quad \theta \in [\theta_{\min},\theta_{\max}],$$2

The description states that this method “very rarely exceeds 5 iterations” and averages about 3 iterations, including component-wise hyperspectral operation (Krivenko et al., 2024).

The feasibility logic is also explicit. If both metric values at the parameter bounds lie below the target, the requested quality is infeasible. If the low-compression boundary already meets or exceeds the target, the tool moves toward higher compression until the result remains within the prescribed tolerance. This makes the system a bounded one-dimensional controller operating on a monotone rate–quality curve rather than a general multidimensional optimizer.

5. Demonstrated operating points and practical use

The paper reports several concrete use cases spanning hyperspectral, aerial, natural, and medical imagery. These examples are used to demonstrate that the tool can converge rapidly to the required metric while reporting the corresponding compression ratio and coder parameter (Krivenko et al., 2024).

Data	Coder and target	Reported result
AVIRIS Lunar Lake hyperspectral sub-band	AGU-M, PSNR-HVS-M = 40 dB	CR = 12.64, QS = 12.43, 3 iterations
Airfield grayscale aerial photo	AGU-M, PSNR-HVS-M = 40 dB	CR = 8.48, QS = 11.33, 2 iterations
Pole grayscale picture	ADCT-M, MSSIM = 0.985	CR = 41.38, QS = 28.9, 4 iterations
MRT medical MRI	ADCT-M, PSNR-HVS-M = 40 dB	CR = 41.41, QS = 19.28, 3 iterations

The practical guidance begins with loading the grayscale image, optionally converting format or applying a homomorphic transform for uneven dynamic ranges, then selecting a coder, metric, target quality, and tolerance. For DCT-based coders, the user sets QS_min and QS_max; for SPIHT and JPEG2000, min/max bpp are set, with typical starting ranges such as $$\min_\theta R(\theta) \quad \text{subject to} \quad Q(I,\hat I(\theta)) \ge q_0 \quad \text{and} \quad \theta \in [\theta_{\min},\theta_{\max}],$$3 noted for high-to-moderate compression ratios (Krivenko et al., 2024).

The examples also motivate the recommendation of AGU-M and ADCT-M when targeting visually grounded metrics. In the reported cases, these HVS-oriented coders achieve compression ratios 30–60% larger than JPEG or JPEG2000 at the same visual quality target. The article does not claim this as a universal bound; it is stated specifically for the reported cases (Krivenko et al., 2024).

6. Limitations, recommendations, and terminological ambiguity

Several limitations are explicitly identified. The tool is focused on grayscale imagery, and color handling is not described. Runtime is not numerically quantified and depends on coder complexity and image size. Computational cost reflects multiple compress–decompress iterations, although the average number of iterations is reported as approximately 3 and rarely more than 5. Artifact characteristics depend on the selected coder, and some target metrics may be infeasible under aggressive compression or poorly chosen search bounds (Krivenko et al., 2024).

The practical recommendations are correspondingly conservative. For visually lossless grayscale optical images, the stated defaults are PSNR-HVS-M = 40 dB or MSSIM = 0.985, with quality accuracy $$\min_\theta R(\theta) \quad \text{subject to} \quad Q(I,\hat I(\theta)) \ge q_0 \quad \text{and} \quad \theta \in [\theta_{\min},\theta_{\max}],$$4 dB for dB-valued metrics or a small tolerance for MSSIM. ADCT-M or AGU-M are preferred when the goal is highest compression ratio at a visual-quality target, while SPIHT and JPEG2000 are preferred when precise bitrate control in bpp is required. JPEG remains the simplest and most widely available option, but may produce more block artifacts at high compression ratios than AGU or ADCT variants (Krivenko et al., 2024).

A common source of confusion is that “How2Compress” is also the name of a later framework for edge video analytics that assigns macroblock-level emphasis maps to H.264 compression. That system is a self-supervised, fine-grained adaptive video-compression module for machine-centric analytics rather than a grayscale image tool for guaranteed metric-targeted compression (Wu et al., 21 Oct 2025). This suggests that the term is polysemous in the recent compression literature and should be interpreted from context.

Within the context of optical-image compression, however, How2Compress denotes a guaranteed-quality control environment: a front end that couples selectable coders, selectable quality metrics, feasibility estimation, and monotone iterative parameter search to deliver a requested output quality together with the maximal attainable compression under that constraint (Krivenko et al., 2024).