Multi-Class Focus Level Classification for Robotic Microscope Using Image Metrics and Machine Learning

Precise focus adjustment is crucial for accurate diagnosis of tuberculosis (TB) using microscopy. Manual focusing remains subjective and labor-intensive, leading to variability in diagnostic outcomes. This study proposes a multi-class focuslevel classification method for robotic microscope autofocus using image-based metrics and machine learning. A dataset of 50 fields of view (FOVs) was collected, covering blur-to-focus transitions at multiple focal depths. Three focus metrics - Tenengrad, Variance of Laplacian, and Normalized Gray-Level Variance were extracted to represent sharpness, gradient, and statistical focus characteristics. The focus levels were classified into five discrete classes based on the normalized Tenengrad values. Three machine learning classifiers were evaluated: Random Forest, KNearest Neighbors, and Support Vector Machine. Among these, the Random Forest model achieved the highest accuracy of 96.22% in the final test set. The proposed approach provides an accurate and efficient focus-level classification and offers strong potential for integration into robotic microscopes for TB diagnostic support.