Cross-Module Interference in QR Codes

• Cross-Module Interference (CMI) is the spill-over of colorant between adjacent high-density QR code modules due to physical diffusion and encoding limitations.
• It is modeled using a linear mixture approach where a central module’s true color is blended with its neighbors, enabling precise parameter estimation for decontamination.
• CMI mitigation, integrated into QR code decoding pipelines, reduces bit-error rates and decoding failures while enabling real-time mobile processing.

Cross-Module Interference (CMI) denotes the contamination of a spatially localized unit’s signal or measurement by neighboring units, arising fundamentally from physical or encoding limitations in densely packed modular structures. CMI has been rigorously analyzed in the context of high-density color QR codes, where it manifests as colorant spill-over between adjacent modules, and is closely related to mutual interference phenomena in other code and waveform design domains. In QR code systems, CMI constitutes a principal source of chromatic distortion, impairing reliable data transmission and decoding fidelity, and requiring explicit modeling and mitigation for high-capacity, robust operation (Yang et al., 2017).

1. Physical Origin and Phenomenology

CMI in high-density color QR codes arises due to the proximity of colored modules (cells), whereby the printed colorant (typically CMY or RGB inks) from one module diffuses or “spills” over into its neighbors during the printing process, and to a lesser extent, during the acquisition phase (scanning/capture). This spill-over acts as a spatial convolution, blending a module’s intended color with those of the immediate adjacent modules. The observed RGB color at position $(i,j)$ therefore represents a linear mixture of the central module’s true color and those of its four (or, if considered, eight) immediate neighbors. By contrast, CMI is distinct from cross-channel interference (caused by colorant–camera channel mixing) or global or spatial illumination variation, as it is strictly a local, spatially induced effect.

2. Mathematical Formulation of CMI

CMI is encapsulated via a linear local mixture model over observed color vectors. For a given module at $(i,j)$, let $X^k$ denote the $5\times3$ matrix whose rows are the normalized RGB values of the central module and its four edge-adjacent neighbors. The “pre-CMI” (decontaminated) color $\tilde{x}$ is modeled as a linear combination: $\tilde{x} = X \theta,$ where $\theta \in \mathbb{R}^5$ is a vector of mixture weights (nearly unity for the center, small—often negative—for neighbors). For each color channel $c\in\{R,G,B\}$,

$$\tilde{x}_c = \sum_{m\,\in\,\{\mathrm{ctr},\mathrm{top},\dots\}} \theta_m X_{m,c}.$$

Estimation of $\theta$ is central: it quantifies the spatial interference kernel induced by colorant diffusion.

3. CMI-Aware Classification Approaches

CMI-aware classification integrates the above mixture model into both generative and discriminative learning frameworks.

3.1 Quadratic Discriminant Analysis with CMI (QDA-CMI)

Each color class $(i,j)$0 in an $(i,j)$1-layer QR code (with $(i,j)$2 codewords) is modeled by a Gaussian density over the decontaminated feature $(i,j)$3: $(i,j)$4 The log-likelihood over all labeled training samples is maximized jointly in $(i,j)$5 and $(i,j)$6: $(i,j)$7 Estimation is performed by alternating between closed-form MLE updates of $(i,j)$8, $(i,j)$9 for fixed $X^k$0, and analytic $X^k$1 updates for fixed class statistics.

3.2 Layered SVM with CMI (LSVM-CMI)

In the layered SVM construct, for each of $X^k$2 layers, a binary SVM is trained, each with its own mixture weights $X^k$3. The layer-specific SVM solves: $X^k$4 Alternating convex optimization between SVM parameters and $X^k$5 via projected gradient yields joint estimation of color-decontamination and classification parameters.

Table 1 summarizes the main distinctions:

CMI-Aware Classifier	Feature Representation	Optimization Scheme
QDA-CMI	Gaussian over $X^k$6	Alternating MLE
LSVM-CMI	Layered SVM on $X^k$7	Alternating SVM/P-GD

4. Practical Pipeline and Robustness Measures

CMI-aware decoding is integrated into a broader high-capacity QR code processing pipeline (“HiQ”) that incorporates robust geometric transformation (RGT), illumination normalization, local binarization, spatial bit randomization, and block accumulation:

• Robust Geometric Transformation (RGT): Employs all detected patterns (finder and alignment) to estimate homography $X^k$8 by minimizing $X^k$9 under $5\times3$0, using a weighted least-squares and SVD.
• Illumination Normalization: Global white estimation from code gaps, followed by channel-wise normalization $5\times3$1.
• Local Binarization: Dynamic thresholding per $5\times3$2 block and channel.
• Bit Randomization and Block Accumulation: Uniformly shuffles bits and aggregates RS blocks over multiple frames to mitigate spatially correlated errors and improve recovery rates.

5. Experimental Evaluation and Impact

CMI-aware decoding was systematically evaluated on the CUHK-CQRC dataset (5,390 high-density, $5\times3$3-layer color QR codes, captured under variable lighting and device conditions). Performance metrics include Bit-Error Rate (BER), Decoding Failure Rate (DFR), and mobile decoding speed.

Method	BER (%)	DFR (%)	iPhone 6P Decode Time (s)
Baseline (PCCC + standard GT)	10.7	84	—
HiQ + QDA (no CMI)	4.3	54	—
HiQ + LSVM (no CMI)	4.3	56	—
HiQ + QDA-CMI	3.9	50	—
HiQ + LSVM-CMI	3.2	46	<0.4

Inclusion of CMI modeling reduced BER by up to 25% (relative) and DFR by 14% over non-CMI versions. LSVM-CMI achieved real-time (<0.4 s) decoding for 6kB codes. CMI mitigation also decreased the minimum decodable QR print size by up to 25% (Yang et al., 2017).

6. Broader Significance and Related Interference Phenomena

CMI exemplifies a dominant spatial interference mechanism in any system relying on localized, high-density modular encoding, especially where the encoding channel exhibits spatially diffuse behavior (e.g., ink diffusion, pixel crosstalk). Methodologies for CMI estimation and mitigation—particularly the linear spatial interference model and joint parameter-classifier optimization—have implications for other domains involving spatial (or spatio-chromatic) cross-talk. This includes, by analogy, slow-time coding adapted for cross-interference mitigation in automotive radar waveforms, where multi-channel interactions must be systematically modeled and suppressed for reliable target detection (Bose et al., 2022).

A plausible implication is that further advances in modular data encoding, print technologies, and signal processing may continue enhancing the robustness of dense modular codes, provided detailed physical modeling of local interference is incorporated into both training and real-time decoding stages.

7. Implementation Considerations and Limitations

CMI-aware classification demands augmented training data, capturing local color mixtures under variable environmental and device scenarios. The linear mixture model for five modules (center + 4 neighbors) provides a tractable, empirically effective compromise between expressivity and statistical/data requirements. Estimation steps can be parallelized and, in the case of LSVM-CMI, are computationally lightweight enough for real-time mobile deployment on contemporary hardware. Nonetheless, extreme-density codes or more severe spill-over may require extension to 8-neighborhood modeling, nonlinear mixture models, or spatially adaptive $5\times3$4 estimation.

The CMI-cancellation approaches do not directly address other sources of error (e.g., geometric distortion, severe non-uniform illumination); hence the need for multi-stage, robust pipeline integration. The success of CMI-aware decoding thus relies on a holistic approach encompassing geometric, photometric, and spatial-interference correction in high-capacity modular encoding systems.