Thermal image super-resolution remains a challenging task due to the limited spatial detail captured by infrared sensors. Although RGB-guided methods and domain adaptation techniques have shown promise, they often introduce architectural complexity or require multimodal inputs at inference time. In this study, we investigate the integration of Fourier Domain Adaptation (FDA) as a lightweight preprocessing strategy to improve the performance of the state-of-the-art Dense-Residual-Connected Transformer (DRCT) model for image super-resolution. FDA transfers low-frequency information from grayscale versions of RGB images to thermal images during training, enabling the model to incorporate additional structural information without introducing any additional architectural complexity or adding overhead during inference. Experimental results on the Tufts Face Database demonstrate that the FDA-enhanced model improves PSNR by up to 0.17 dB and SSIM by up to 0.0004 compared to thermal-only training at a x2 scaling factor. These findings highlight the effectiveness of simple frequency-based adaptation techniques in improving the generalization of thermal SR models while preserving the simplicity of the inference pipeline.

Camera-based vital sign estimation is a very active area of research for human-system interaction and medical diagnostics. The sequence of frames is processed to estimate respiration rate, blood volume pulse (BVP), emotions, etc. In remote photoplethysmography, the source of the blood pulse signal is the arteries in the skin. This study aims to investigate facial skin segmentation methods for rPPG signal estimation. Three methods were adapted and evaluated: a threshold-based method (YCrCb/HSV), deep semantic segmentation using the BiSeNet model, and an attention-based approach derived from Grad-CAM visualizations of the TS-CAN network. Each method was used to generate a facial mask applied to extract the green-channel intensity signal from video frames. Signal quality was evaluated using two metrics:$\mathbf{S N R}_{{raw }}$and$\mathbf{S N R}_{max}$. Additionally, IoU metrics were used to quantify spatial overlap between the masks. The results show that BiSeNet segmentation yielded the highest signal quality ($\mathbf{S N R}_{max }=\mathbf{7. 8 1 8} \mathbf{dB}$), while Grad-CAM masks, applied without additional processing, achieved the lowest performance ($\mathbf{S N R}_{max }=-\mathbf{1. 5 3 5} \mathbf{~ d B}$). This indicates that although attention maps can identify important facial regions, further refinement is required to improve their effectiveness in$r$PPG signal extraction. Spatial analysis revealed strong agreement between the threshold-based and BiSeNet masks, with significantly lower alignment between either and the Grad-CAM-based mask. An additional analysis of Grad-CAM attention distribution across facial regions showed that the model focused mainly on the cheeks, regions associated with more stable pulsatile signals. These findings emphasize the importance of accurate skin segmentation for enhancing rPPG signal quality and show that attention maps used for interpreting network behavior do not precisely reflect the regions relevant for accurate skin segmentation. This work supports the development of more explainable and reliable data-driven rPPG systems.

Recent inpainting methods have demonstrated im-pressive outcomes in filling missing parts of images, especially for reconstructing facial areas obscured by occlusions. However, studies show that these models are not adequately effective in real-world applications, primarily due to data bias and the distribution of faces in images. This research focuses on domain adaptation of the commonly used Labeled Faces in the Wild (LFW) dataset, employing the Mask-Aware Transformer (MAT) inpainting method for reconstructing occluded facial regions and examining its impact on facial recognition accuracy. Three types of generated masks were applied to specific facial areas, covering key points on the face, using three datasets: CelebA-HQ, LFW, and a specially adapted LFW. The analysis employed various metrics to assess the quality of the reconstruction. The results indicate that applying a simple adaptation method to the LFW dataset significantly boosts facial recognition capabilities, with improvements reaching up to 16.43% compared to the original LFW. Subsequently, the experiments demonstrate that using inpainting methods enhances face recognition considerably when compared to images with applied masks without reconstruction. Notably, improvements in positively verified images were ob-served up to 89.30% for CelebA-HQ, 21.37% for LFW, and 29.69% for the adapted LFW.

Pregnancy in a life of a woman, is an important time that is connected with both physiological and psychological changes. This paper aims at developing a digital twin application that allows to assess mother's health risk and help to diagnose them. The system presented in this paper includes models for three health outcomes: maternal health risk level, diagnosis of gestational diabetes mellitus (GDM), and diagnosis of late onset preeclampsia. The system included an examination of a data generation method. The model destined to assess the risk level achieved an accuracy of 83.5%. GDM model obtained a high precision of 97.2%. The analysis of preeclampsia data generation has shown a great potential for future use. The developed digital twin application serves to exhibit the mentioned models and offer an insight into the future diagnostic tool designed for maternal healthcare in the coming years.