RGB-Guided Infrared Fixed-Pattern Noise Estimation via Sequential Adaptive Fusion
IEEE Sensors Journal, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/jsen.2026.3707383
- Dergi Adı: IEEE Sensors Journal
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Compendex, INSPEC, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Anahtar Kelimeler: Deep learning, fixed pattern noise, infrared image denoising, non-uniformity correction, recurrent neural networks, rgb guidance, sensor noise modeling
- Hacettepe Üniversitesi Adresli: Evet
Özet
Fixed-pattern noise (FPN) is a persistent artifact in infrared (IR) imaging systems, arising from sensor non-uniformities and optical inconsistencies. In this work, we introduce a deep learning-based architecture, SAFTA (Sequential Adaptive Fusion for Thermal Artifact Removal), that jointly estimates clean infrared images and the underlying noise parameters by leveraging paired RGB and IR observations, and by explicitly estimating the gain and offset parameters responsible for FPN. The proposed framework consists of two main components: an RGB-guided infrared estimator and a recurrent FPN parameter estimator. The first stage employs an encoder–decoder network to predict a clean infrared image using spatial cues from both RGB and noisy IR inputs. The second stage refines this prediction by estimating the gain and offset noise parameters through a gated recurrent unit (GRU)-based recurrent module, which iteratively processes a sequence of image pairs and updates the FPN representation over time, effectively capturing temporally consistent noise patterns. The combination of spatial fusion and temporal noise adaptation enables robust correction of both high and low-frequency non-uniformities in thermal images. Unlike conventional approaches that process each frame independently, our method reuses noise parameters across frames, leading to improved temporal stability and enabling fast and consistent correction suitable for real-time applications. The proposed method is evaluated on both real-world and synthetically generated FPN datasets and outperforms recent state-of-the-art approaches across multiple image quality metrics.