Image Demoiréing via Multiscale Fusion Networks With Moiré Data Augmentation

In digital imaging, Moiré patterns pose a significant challenge, appearing as rainbow-colored, warped grid patterns. These artifacts emerge when two overlapping patterns create color distortions due to their slight differences. The limitations of complementary metal-oxide-semiconductor (CMOS) active pixel sensors become apparent in photographs, as these sensors capture information discretely in pixels. This is particularly pronounced when photographing patterned items, mainly due to misalignment issues between the sensor and the regularly spaced, overlapping stripes. This process creates a disparity between real-world continuous lines and digital screens, thereby complicating the work of professional photographers. In response to this challenge, we propose the demoiréing multiscale fusion network (DMSFN). By leveraging dilated-dense attention (DDA), multiscale feature interaction, and multikernel strip pooling (MKSP), our approach effectively detects and removes Moiré patterns from sensor outputs. To enhance the demoiréing performance, we augment the training data by transferring Moiré patterns to clean images. Our experimental results indicate that our proposed model outperforms existing state-of-the-art (SOTA) demoiréing methods, as validated on benchmark datasets. This work contributes to advancing the quality of digital sensor outputs in the presence of Moiré patterns and addressing challenges encountered in practical applications.