CIPPSRNet: A Camera Internal Parameters Perception Network Based Contrastive Learning for Thermal Image Super-Resolution

Kai Wang, Qigong Sun, Yicheng Wang, Huiyuan Wei, Chonghua Lv, Xiaolin Tian, Xu Liu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 342-349

Thermal Image Super-Resolution (TISR) is a technique for converting Low-Resolution (LR) thermal images to High-Resolution (HR) thermal images. This technique has recently become a research hotspot due to its ability to reduce sensor costs and improve visual perception. However, current research does not provide an effective solution for multi-sensor data training, possibly driven by pixel mismatch and simple degradation setting issues. In this paper, we proposed a Camera Internal Parameters Perception Network (CIPPSRNet) for LR thermal image enhancement. The camera internal parameters (CIP) were explicitly modeled as a feature representation, the LR features were transformed into the intermediate domain containing the internal parameters information by perceiving CIP representation. The mapping between the intermediate domain and the spatial domain of the HR features was learned by CIPPSRNet. In addition, we introduced contrastive learning to optimize the pretrained Camera Internal Parameters Representation Network and the feature encoders. Our proposed network is capable of achieving a more efficient transformation from the LR to the HR domains. Additionally, the use of contrastive learning can improve the network's adaptability to misalignment data with insufficient pixel matching and its robustness. Experiments on PBVS2022 TISR Dataset show that our network has achieved state-of-the-art performance for the Thermal SR task.