Infrared imagery is crucial for surveillance, object detection, and autonomous navigation applications. However, the inherent low resolution of IR sensors limits the extraction of fine-grained details necessary for downstream tasks. Previous super-resolution approaches have predominantly optimized performance for IR single-image super-resolution; however, existing methods overlook spatial error analysis and real-world validation. This article presents an Edge Enhanced Deep Super-Resolution Network specifically designed for IR images by integrating a three-component loss function; pixel-wise L1 loss, edge-preserving loss, and perceptual loss. EEDSR is specifically tailored to ensure color fidelity, sharp edge reconstruction, and perceptually realistic outputs, unlike prior methods. Experiments performed on benchmark datasets and real-world captured IR images demonstrate significant improvements over existing methods, achieving a PSNR of 30.47 and an SSIM of 0.8810. Furthermore, comprehensive error map analysis highlights superior structural preservation and reduced spatial artifacts. The proposed approach bridges the gap between high-fidelity IR reconstruction and practical applicability, paving the way for more vigorous IR-based vision systems. Suárez, Patricia L., Dario Carpio, and Angel D. Sappa. “Enhancement of Guided Thermal Image Super-Resolution Approaches.” Neurocomputing 573 (March 2024): 127197. https://doi.org/10.1016/j.neucom.2023.127197.
We propose SepScene, a proxy-guided framework implementing inherently separable 3D scenes with native per-component manipulation. Using semantically annotated proxy scenes, our approach separates scene elements based on movability attributes, i.e., distinguishing immovables such as walls and terrain from user-relocatable movables including furniture and props. Environment-anchored immovables undergo geometry-constrained depth prediction to ensure structural permanence, while user-relocatable movables are processed through occlusion-aware diffusion inpainting followed by single-view mesh reconstruction, preserving their inherent reconfigurability. Critical challenges including occlusion handling and semantic consistency are resolved through proxy-guided raster masks and ID-mapped text embeddings, with object details refined via ratio-adaptive hallucination. The SepScene system establishes substantial post-generation control by enabling direct transformation, replacement, or removal of scene components without full regeneration. Our method surpasses existing approaches by achieving superior scene-text alignment and layout plausibility while also offering enhanced perceptual quality and faster generation speed.
