STRENGTHENING LOCAL SELF GOVERNMENT THROUGH ACCOUNTABLE DIGITAL PUBLIC SERVICES AND VISUAL DATA QUALITY MANAGEMENT A FRAMEWORK FOR MOIRÉ BOUNDARY DETECTION

Povzetek

Moiré patterns, caused by interference between repetitive structures and sensor sampling frequencies, pose a persistent challenge in digital imaging and computer vision tasks. These artifacts degrade visual quality and significantly hinder downstream processes such as image restoration and enhancement. While prior research has introduced deep learning frameworks for moiré pattern detection and removal, many approaches struggle with boundary inconsistencies and limited generalization across diverse image conditions. To address these challenges, this paper proposes a novel framework for accurate moiré boundary detection by combining multi-source triplet training and edge context fusion. The model leverages three aligned data types—clean natural images, isolated moiré layers, and synthetic mixtures—to ensure reliable pixel-level correspondence and strong discriminative learning capability. The proposed architecture integrates a multi-scale encoder for local and global feature extraction with an adaptive fusion mechanism that jointly analyses spatial texture and frequency signals. This design enables precise delineation of moiré interference boundaries while maintaining low computational cost. Experiments conducted on public datasets, including MoireScape and MoireIDT, demonstrate that our approach yields superior boundary localization and robustness compared to existing baselines. The method not only advances moiré pattern boundary detection but also provides a strong foundation for subsequent demoiréing and restoration networks, improving real-world camera image quality analysis and enhancement workflows.