First set of facial skin images and decision of methods for preprocessing

Image Preprocessing Algorithms The preprocessing pipeline (described in Deliverable D2.1) combines several steps designed to ensure the spatial and temporal consistency of facial sequences before their analysis through the STREAM-Net model. First, the RGB images are resized to a fixed resolution of 144 × 144 px, and face cropping (ROI extraction) is applied to remove background regions and focus the analysis on skin areas relevant to blood flow estimation. Then, a temporal normalization between consecutive frames is performed, computed as the difference divided by the sum of consecutive frames, in order to highlight dynamic color variations associated with blood pulsation and reduce the influence of illumination or skin tone. This processing generates two complementary inputs —a spatial and a temporal one— that feed the dual-branch architecture of STREAM-Net. Additionally, a Bayesian uncertainty estimation mechanism using Monte Carlo Dropout is integrated to assess the model’s robustness to stochastic variations and quantify prediction confidence. Finally, to further improve image quality and facilitate more accurate physiological signal extraction, super-resolution and deblurring techniques based on deep learning (e.g., EDSR, LapSRN, SwinIR, and FaceSR) are applied, enhancing both fine-detail recovery and structural fidelity according to PSNR and SSIM metrics. PURE Dataset The PURE (Pulse Rate Detection) dataset provides facial video sequences used for the training and validation of remote pulse estimation algorithms such as those implemented in STREAM-Net. It includes recordings of 10 volunteers (8 male, 2 female) performing six controlled motion conditions —steady, talking, slow translation, fast translation, small rotation, and medium rotation— introducing different levels of motion and illumination variability. Each sequence lasts one minute and was recorded using an eco274CVGE camera (SVS-Vistek GmbH) at 30 fps and 640 × 480 px resolution, while reference pulse signals were simultaneously acquired using a Pulox CMS50E oximeter at 60 Hz. The participants were seated approximately 1.1 m from the camera under natural daylight illumination, with slight variations caused by cloud movement. These recordings capture subtle color changes on the skin surface as well as realistic head movements, providing an ideal dataset for evaluating remote photoplethysmography (rPPG) pipelines and assessing the robustness of preprocessing algorithms against motion artifacts, illumination fluctuations, and facial expressiveness.