Cuffless blood pressure (BP) detection is beneficial for improving the diagnosis and treatment of hypertension. Current intermittent blood pressure measurement methods neither fully capture an individual’s true blood pressure status nor completely eliminate the constraints of the cuff. This study introduces a novel model, high and low frequency information fusion network (HLNet). This approach combines an innovative patch mixing design to reveal the inherent complex temporal patterns in photoplethysmography (PPG) sequences. Additionally, we introduce a convolutional operator called LocAttn, which incorporates self-attention mechanism that effectively utilizes shared and contextual soft weights for local perception. LocAttn employs a novel approach compared to traditional local self-attention, incorporating stronger nonlinearities to generate Context-adaptive weights. Furthermore, a dual-branch structure is utilized, with one branch employing LocAttn to capture high-frequency information, and the other branch utilizing downsampled periodic attention to capture low-frequency information. This dual-branch architecture enables HLNet to simultaneously fuse high-frequency local information and low-frequency global information. On a subset of the MIMIC dataset, the model’s detected systolic blood pressure (SBP) and diastolic blood pressure (DBP) errors meet The Association for the Advancement of Medical Instrumentation (AAMI) standard. The proposed method achieves optimal error performance with a mean error of 0.06 ± 4.41 mmHg for SBP and 0.07 ± 2.62 mmHg for DBP (ME ± SD). The SBP and DBP accuracy meets the British Hypertension Society (BHS) standard of grade A.