CMILK: correlation-based multiple imputation with local k-neighbour matching for missing thermal facial data

Thermal facial images are used to measure body temperature, and facial landmarks are introduced in computer vision tasks to aid in feature extraction. These landmarks help extract temperature data from thermal facial images. However, landmarks may be absent either due to specific factors such as head pose variations or occlusions from glasses or facial accessories, this is classified as Missing Not at Random (MNAR), or they may be absent without any systematic cause, which is referred to as Missing Completely at Random (MCAR). To overcome this significant issue, we proposed a method called CMILK: Correlation-based Missing Landmark Imputation using Local k-neighbours, which combines Pearson’s Correlation Coefficient (PCC) with local k-neighbour matching to predict missing temperature values accurately. Comparative evaluations show that CMILK achieves a 5% improvement in Root Mean Square Error (RMSE), a marginal enhancement in Mean Absolute Error (MAE), and a substantial reduction in computation time, up to 13 times faster than the next best-performing method. The substantial gain in RMSE demonstrates that CMILK is more robust to outliers and large errors. The proposed method advances accurate data imputation methodologies, enhancing reliability in predictive modelling and robust analysis of thermal facial landmark temperature data.