How to Detect and Remove Moiré Patterns with AI and Diffusion Models

Introduction

Moiré patterns are high‑frequency interference stripes that appear on images captured by digital cameras or scanners when the sensor grid interferes with fine details of the scene. While they can be used creatively for depth, they usually degrade visual quality.

Manual Mitigation Methods

Adjust camera angle to reduce interference.

Change focus to avoid overly sharp details.

Vary lens focal length.

Use software (e.g., Photoshop) to tweak contrast, brightness, and color.

Apply low‑pass filters to suppress high‑frequency signals.

Moiré Pattern Recognition

Accurate detection is critical for Huolala’s business scenarios such as moving‑in check‑in, driver‑license verification, and vehicle‑sticker inspection.

Direct Methods

These classify images as moiré or normal using either traditional image processing or deep learning.

Traditional Methods

Spatial‑domain features: Edge detection (Sobel, Laplacian) to locate visible lines.

Frequency‑domain features: Fourier or wavelet transforms reveal characteristic peaks.

Statistical features: Histogram analysis combined with Local Binary Patterns (LBP).

Deep Learning Methods

A convolutional neural network (CNN) extracts hierarchical features and classifies images after training on moiré and normal samples.

In practice, traditional preprocessing (edge, frequency, or statistical features) is often combined with a CNN for final classification.

Indirect Methods

These first apply a demoiré model to the input image and then compare the before‑and‑after results to decide whether the original contained moiré.

Demoíré Models

Recent models such as MBCNN, WDNet, and ESDNet are discussed, with a focus on ESDNet’s encoder‑decoder architecture and its SAM (Scale‑Aware Module) that fuses multi‑scale features via dynamic weighting.

Image Quality Evaluation Metrics

PSNR : Higher values indicate closer similarity; moiré removal reduces similarity for moiré images.

SSIM : Measures structural similarity; drops more for moiré images after demoiré.

MSE : Larger for moiré images after processing.

CIEDE : Color difference metric.

Weighted combination of the above for a composite score.

High‑Frequency Information Statistics

High‑frequency band statistics decrease markedly after demoiré for moiré images but not for normal images.

Difference of high‑frequency components via Fourier analysis.

Frequency‑domain morphological processing to detect disappearing large connected components.

Information Entropy Difference

Convert images to grayscale, compute normalized histograms, then calculate entropy. Moiré images typically show a negative entropy change after demoiré, while normal images show a slight increase.

Practical Deployment at Huolala

The moiré detection pipeline is used in moving‑in check‑in, driver‑license verification, and vehicle‑sticker review to automatically flag forged or duplicated photos, reducing manual workload and improving accuracy.

Moving‑In Check‑In Scenario

Drivers submit a photo before packing; the system flags images with moiré patterns as potential forgeries, ensuring authenticity and preventing fraud.

Diffusion Model for Moiré Detection

Inspired by recent advances, Huolala adopts the DIAD diffusion framework. During training, only normal images are used. In inference, the model receives both a noisy (potentially moiré) image and the original image; the SG module reconstructs the moiré‑free version. If the output differs from the input, the image is classified as moiré.

Summary and Outlook

Achieving high recall and low false‑positive rates remains the main challenge. Diffusion models provide a promising solution, and future work will focus on further algorithmic refinements and exploring new techniques to boost performance.

References

Zhang, Bolun, et al. "Image Demoireing with Learnable Bandpass Filters." CVPR 2020.

Liu, Lin, et al. "Wavelet‑based and Dual‑branch Neural Network for Demoireing." ECCV 2020.

Yu, Xin, et al. "Towards Efficient and Scale‑Robust Ultra‑HD Image Demoireing." ECCV 2022.

Zhang, Yuxin, et al. "Real‑Time Image Demoireing on Mobile Devices." ICLR 2023.

Sohl‑Dickstein, Jascha, et al. "Deep Unsupervised Learning using Nonequilibrium Thermodynamics." ICML 2015.

Ho, Jonathan, et al. "Denoising Diffusion Probabilistic Models." NeurIPS 2020.

Rombach, Robin, et al. "High‑Resolution Image Synthesis with Latent Diffusion Models." CVPR 2022.

He, Haoyang, et al. "DiAD: A Diffusion‑based Framework for Multi‑class Anomaly Detection." AAAI 2024.