Image-based photoplethysmography (IPPG) holds promise for applications like health surveillance and emotional state analysis. Despite recent progress in crafting deep learning-centric IPPG methodologies, which predominantly forge a correlation between spatiotemporal heart rate (HR) feature imagery and corresponding HR readings, these techniques encounter constraints in extended spatiotemporal comprehension and engagement. In this manuscript, we introduce the BiFormer architecture, an end-to-end solution integrating temporal difference convolution, multi-head self-attention transformer modules, and bidirectional long short-term memory networks to refine signal estimations and bolster the model’s discernment prowess. Our framework was appraised through intra-database and inter-database evaluations on three accessible datasets, evidencing superiority over conventional IPPG strategies in HR accuracy metrics. Notably, assessments on the VIPL-HR dataset indicated a reduction in the average root mean square error to 7.24 beats per minute.