Machine Learning-Based Computer Vision Optimization for Marker Detection in Augmented Reality Mathematics Learning

Marker-based Augmented Reality (AR) applications for mathematics learning remain sensitive to variations in illumination, marker rotation, camera distance, and viewing angle, which can reduce detection accuracy and tracking stability. This study investigates a computer vision optimization pipeline integrating Contrast Limited Adaptive Histogram Equalization (CLAHE), Oriented FAST and Rotated BRIEF (ORB), Support Vector Machine (SVM), and Random Sample Consensus (RANSAC) to improve marker detection and tracking performance. The proposed pipeline enhances local image contrast, extracts visual features, classifies marker descriptors, and rejects inconsistent feature correspondences during homography estimation. A quantitative experimental evaluation was conducted under normal and low illumination, marker rotation, and varying camera distances, with performance assessed using classification metrics, detection time, frame rate, and tracking stability. The reported results indicate an accuracy of 97.82%, precision of 97.35%, recall of 98.10%, and F1-score of 97.72%. The system also achieved an average detection time of 31.6 ms, a processing rate of 30.8 frames per second (FPS), and tracking stability of 98.41% under the reported evaluation conditions. Compared with the stated ORB-based baseline, which achieved 91.84% accuracy, 38.72 ms detection time, 24.60 FPS, and 90.74% tracking stability, the proposed pipeline demonstrates improved reported detection and tracking performance. These findings suggest that combining local contrast enhancement, feature-based classification, and geometric outlier rejection is a promising approach for marker-based AR mathematics learning. Further validation using explicitly documented hardware configurations, reproducible algorithm parameters, and expanded environmental stress tests is needed to establish the generalizability of the results.