We introduce an arterial blood pressure reconstruction method from photoplethysmography signals obtained by a cuff-less device. The proposed method is built upon a UNet-based architecture with the difference of Gaussian (DoG) attention module. With multi-scale features from the encoder of the UNet, the DoG attention block emphasizes important sub-band frequency parts of features, ultimately leading to improved blood pressure estimation performance. Utilizing the MIMIC-III dataset, which comprises over 12,000 records, the proposed method shows better performance than the existing methods in the Association for British Hypertension Society standard and reasonable performance in the Advancement of Medical Instrumentation standard. In the future, we hope that deploying the proposed method in a sensor robot or a wearable device can lead to widespread use in personal health management and remote patient monitoring scenarios.