Infrared Camera Array Image Super-Resolution based on Zero-Shot Learning

The infrared camera array is capable of capturing infrared features in a single scene from multiple views, thereby supplying rich angular information for a range of post-capture tasks, one of which is the image super-resolution (SR). Since the spatial resolution of each sub-image is limited by the diffraction effects, prevalent SR methodologies endeavor to improve image quality by acquiring prior knowledge from extensive datasets. However, the scarcity of the publicly available infrared array image dataset makes it difficult to rely on conventional model training strategies. To address this, inspired by the recent advance in zero-shot learning and light field image SR, this paper proposes a zero-shot infrared array image SR method. Specifically, we design an array feature extraction module (AFEM) to incorporate latent correlations within views and employ a zero-shot training strategy to enable model learning with one input low-resolution scene. Experiments show that the proposed method can achieve reliable performance in both objective metrics and visual effects.