Stability-Aware Video-Window-Wise Bayesian Optimization of CLAHE for Ordinal Drowsiness Classification in Near-Infrared Facial Video

Abstract

Contrast Limited Adaptive Histogram Equalization (CLAHE) can improve near-infrared (NIR) facial video, but the temporal granularity at which its parameters are optimized may affect both visual consistency and downstream sequence learning. This study reformulates CLAHE tuning as stability-aware video-window-wise Bayesian Optimization (BO), where one parameter pair is applied to all frames in a ten-second input window and the objective jointly represents perceptual quality and temporal consistency. The analysis uses 2,140 DROZY facial-video windows from 14 participants and 36 usable sessions. Representation-level experiments show a systematic quality–stability trade-off: the quality-only objective attains the lowest normalized PIQE(0.3228), whereas SA 30:70 attains the highest tSSIM (0.7928) and lowest FDV (107.84). The complete eight-condition classifier experiment was evaluated using strict leave-one-subject-out (LOSO) validation, producing 112 separate MobileNetV2–GRU LOSO training runs. Under pooled out-of-subject evaluation, SA 30:70 achieved the highest Accuracy (0.5949; 95% subject-cluster bootstrap CI 0.4976–0.6935) and QWK (0.3147; 0.1337–0.4897), Flatten/legacy Window-wise BO achieved the highest Macro-F1 (0.4580; 0.3675–0.5517), and Static CLAHE remained highly competitive (Accuracy 0.5921; Macro-F1 0.4375; QWK 0.3050). Friedman tests indicated overall differences for subject-level Accuracy (p = 0.0138) and fixed-three-class Macro-F1 (p =0.0348), but no pairwise comparison survived Holm correction. Paired two-one-sided equivalence tests showed that SA 30:70 is equivalent to Static CLAHE and to legacy Window-wise BO within ±0.05 subject-level Accuracy and Macro-F1, while providing markedly higher temporal stability (tSSIM 0.793 versus 0.721; FDV 108 versus 311 relative to Static CLAHE).  The pattern was retained with training-fold class weighting, which raised Neutral recall by 0.14–0.20 and made every leading CLAHE condition significantly better than Original after Holm correction, and with a MobileNetV2–LSTM classifier, where SA 30:70 obtained the highest pooled Accuracy (0.6112). Under an equal 13-evaluation budget, BO outperformed Random Search on 85–91% of windows and matched a structured grid. Stability-aware optimization required approximately 5.9 s per ten-second window on one CPU thread, versus 16.9 s for frame-wise BO. The results support temporal stability as a preprocessing design dimension that can be raised substantially without reducing unseen-subject classification beyond the equivalence margin, while not establishing a universally superior CLAHE configuration.