Infrared thermal image denoising with symmetric multi-scale sampling network

We propose an infrared thermal image denoising method based on the residual learning with a symmetric multiscale (SM) encoder-decoder sampling structure (SMEDS). The U-shape-based SMEDS is designed to extract SM information from different layers and focus on decoding recovery with attention acting on the upper and lower levels. Specifically, SMEDS consists of SM encoder-decoder blocks, cascaded residual blocks, and attention recovery modules. An attention-guided reconstruction unit (AGRU) and SSIM loss are jointly used to supervise image reconstruction and enhance visual perception. A novel Gaussian-Poisson-Stripe noise (GPSN) model is developed to simulate real-world noise. To verify the effectiveness of SMEDS, ablation studies are conducted. Extensive experiments demonstrate that the proposed method performs well on synthetic and real-world noisy images and outperforms previously reported infrared image denoising methods. The proposed method has great potential applications in areas such as remote sensing, infrared medical imaging, and navigation.