Researchers' structural guidance framework sharpens camera image reconstruction

Joint demosaicing and denoising is a critical first step in any camera's image signal processing (ISP), yet it remains stubbornly hard. Raw sensor data is both color-sampled (through a Bayer-like CFA pattern) and noisy, which corrupts edges, repetitive textures, and fine patterns such as moiré. Existing unified restoration networks explicitly model CFA geometry but rely mainly on pixel-level supervision, so they often get structural details wrong where local evidence is ambiguous.

To fix this, Qixin Zheng and collaborators propose a structural-guided unified restoration framework. The model receives a unified five-channel input: the raw mosaic, CFA masks, and a noise-level map. A SwinIR restoration branch reconstructs pixel details under CFA-conditioned modulation, while a parallel structural reasoning branch extracts complementary cues from a sparse pseudo-RGB observation. Since raw sensor data lives far from natural-image pretraining, a lightweight trainable adapter bridges the domain gap before residually fusing structural features. A shared decoder predicts both the restored RGB image and an auxiliary clean mosaic, supervising in both image and sensor domains. Across multiple CFA patterns and noise levels, the approach consistently beats state-of-the-art unified and CFA-specific methods, showing that adapted structural priors genuinely improve robust camera image restoration.