This study presents a deep learning framework to suppress motion artifacts in smartphone-based cPPG. We developed a customized 1D U-Net combined with a hybrid CNN-LSTM network, utilizing physiological data augmentation to address data scarcity. The model achieved a Pearson correlation of 0.848 and a heart rate MAE of 0.186 bpm, effectively preserving essential waveform morphology such as the dicrotic notch. With only 633 K parameters (2.41 MB), this lightweight architecture enables precise, real-time cardiovascular monitoring on mobile edge devices.