A Two-Stage Cascaded Regression Framework for Automatic Facial Acupoint Localization in Infrared Thermal Images

Infrared thermal imaging offers objective physiological insights for Traditional Chinese Medicine (TCM), yet automated acupoint localization struggles with low texture and extreme pose variations. To address this, we constructed a multi-pose thermal facial dataset using digitized bone-proportional measurements and TCM anatomical rules. We propose T2FAL, a two-stage cascaded regression framework explicitly decoupling macroscopic face detection from fine-grained acupoint localization. Stage 1 utilizes an improved YOLOv12m-based Thermal-Aware Face Detector—integrating ICAN_C2f, MixNeck, and TAF-IoU—to mitigate domain shifts and thermal noise. Stage 2 deploys pose-specific regressors incorporating Gated Feature-Conditioned Cascade Refinement (FCCR) and a Selective GeoDeriv module, mathematically translating anatomical rules into geometric constraints. Under the stated experimental protocol, Stage 1 processed images at 87.3 frames per second on an NVIDIA RTX 4090. For Stage 2, the final frontal- and profile-view configurations achieved mAP@50–95 values of 73.19% and 86.92%, with mean pixel errors of 1.986 and 3.109 pixels, respectively. These results support the feasibility of automated reference-coordinate localization on the datasets used. However, given the partial reliance of the annotations on image registration and geometric rules, the use of the cross-domain set for model selection, and the lack of clinical or diagnostic evaluation, further validation using independently generated expert annotations, a strictly held-out external test set, and clinically labeled data is required before clinical or diagnostic use.