Per-Image PCA and Inter-Channel Redundancy Analysis of the Kodak Lossless True Color Image Suite

Baetzel, J. (2026)

This repository contains the first published per-image statistical characterization of all 24 images in the Kodak Lossless True Color Image Suite (PCD0992). Each image is documented as a two-page reference data sheet reporting the complete inter-channel redundancy structure: covariance matrix, eigendecomposition, Pearson correlations, spatial autocorrelation, and derived classification metrics.

All statistics were computed directly from the 8-bit RGB pixel arrays of the standard 768x512 base-resolution PNG distribution. No subjective descriptions appear in any profile. All redundancy classifications are generated programmatically from the computed metrics using fixed thresholds documented in the methodology.

Parent Paper: Baetzel, J. (2026). Statistical Characterization of Inter-Channel Redundancy Structure in the Kodak Lossless True Color Image Suite. Per-Image Principal Component Decomposition of PCD0992.

• Focus: Theoretical framework establishing the first complete per-image PCA decomposition of the Kodak suite. Documents the dimensionality spectrum, blue channel independence range, eigenvector loading patterns, and evidence for deliberate curation across the 24-image collection.
• Availability: Included in this repository (baetzel_2026_kodak_pca_characterization.pdf).

This Series: Baetzel, J. (2026). Kodak PCD0992 Statistical Profile Series. Per-Image PCA and Inter-Channel Redundancy Analysis.

• Focus: Individual reference data sheets and machine-readable metric exports for each of the 24 images. Provides the per-image evidence underlying the suite-wide analysis in the parent paper — covariance matrices, eigendecompositions, correlation heatmaps, spatial autocorrelation, and logic-generated redundancy classifications.
• Availability: /baseline/ directory (24 PDFs + 25 JSON files).

The parent paper establishes why the Kodak suite spans the full spectrum of inter-channel redundancy. The profile series documents what each individual image contributes to that spectrum.

Each two-page profile reports the following:

Page 1

• RGB channel distribution (smoothed density curves from pixel data)
• Per-channel statistics: mean, standard deviation, variance, kurtosis, skewness, min, max
• Inter-channel correlation heatmap (3x3)
• Pairwise Pearson correlation coefficients (R-G, R-B, G-B) and suite average
• Full 3x3 covariance matrix

Page 2

• Eigendecomposition: eigenvalues, variance explained (%), eigenvector loadings
• Derived metrics: condition number, eigenvalue ratios, blue channel independence, PC1 dominant channel
• Dimensionality tier classification
• Spatial autocorrelation (lag-1, horizontal and vertical)
• Average local variance (3x3 neighborhood)
• Redundancy profile (logic-generated from computed metrics)

The 24 images span nearly the full range of inter-channel redundancy configurations achievable through film-based photographic capture. Condition numbers range from 7.55 to 1,739.16 — more than two orders of magnitude — covering color distributions from near-spherical to extremely elongated.

Condition Number (lambda1/lambda3): Ratio of the largest to smallest eigenvalue. High values indicate a needle-like color distribution concentrated along one axis. Low values indicate a more spherical distribution where each channel carries independent information.

Blue Channel Independence: The percentage of blue channel variance not captured by the first principal component. Computed as (1 - (blue_loading_PC1^2 x lambda1 / Var(B))) x 100. Low values indicate the blue channel is almost entirely predictable from the primary variance axis. High values indicate the blue channel carries substantial unique information.

Dimensionality Tier: Classification based on PC1 variance explained. Thresholds at 75%, 85%, 93%, and 97% produce five tiers from Three-Dimensional to Near-Degenerate, corresponding to distinct regimes of inter-channel redundancy.

Eigenvector Pattern: The loading structure of the first principal component. Identifies which channel or channel pair drives the dominant variance axis: balanced (all channels near-equal), coupled (two channels co-load), or dominant (one channel leads).

Spatial Autocorrelation (lag-1): Pearson correlation between each pixel and its immediate neighbor, computed separately for horizontal and vertical directions. Values near 1.0 indicate smooth, spatially coherent image data.

/
    README.md
    baetzel_2026_kodak_pca_characterization.pdf
/baseline/
    KODIM01_STATISTICAL_PROFILE.pdf
    kodim01_stats.json
    KODIM02_STATISTICAL_PROFILE.pdf
    kodim02_stats.json
    ...
    KODIM24_STATISTICAL_PROFILE.pdf
    kodim24_stats.json
    kodak_suite_master_stats.json
/docs/
    methodology.md

Root: The parent PCA characterization paper and repository README. /baseline/: 24 two-page PDF reference data sheets and 25 JSON files (24 individual + 1 master). /docs/: Computation pipeline documentation for full reproducibility.

[1] Eastman Kodak Company. Kodak Publication No. PCD-042, 1992.

[2] Baetzel, J. (2026). “Statistical Characterization of Inter-Channel Redundancy Structure in the Kodak Lossless True Color Image Suite.”

[3] Watanabe, S. “Karhunen-Loeve Expansion and Factor Analysis,” pp. 635-660, 1965.

[4] Giorgianni, E.J. and Madden, T.E. Digital Color Management. Addison-Wesley, 1998.

Baetzel, J. (2026). Kodak PCD0992 Statistical Profile Series:
Per-Image PCA and Inter-Channel Redundancy Analysis of the
Kodak Lossless True Color Image Suite.

Statistical analysis and profile sheets by Jasmine Baetzel (2026). Benchmark images from the Kodak Lossless True Color Image Suite (PCD0992), released by Eastman Kodak Company for unrestricted usage.