The Effect of Dataset Count on Facial Recognition Accuracy Using Haar Cascade Classifier

This study investigates the impact of varying facial image dataset sizes on the accuracy of facial recognition using the Haar Cascade Classifier method. The dataset sizes examined were 150, 100, and 50 facial images, all captured under consistent conditions using a Raspberry Pi Camera v1.3. The dataset collection, image training, and facial recognition processes were conducted on a Raspberry Pi 4 Single Board Computer (SBC). The study controlled for lighting conditions to ensure they did not affect the results. The facial images were trained using the Local Binary Pattern Histogram (LBPH) recognizer and the Haar Cascade Classifier detector. Recognition tests were conducted at distances of 25cm, 50cm, 75cm, 100cm, 125cm, and 150cm from the camera. The study found that the highest recognition accuracy of 98% was achieved with a 150-image dataset at a distance of 75cm. In contrast, the accuracy dropped to 0% at 125cm and 150cm distances across all dataset sizes. The findings suggest that increasing the number of facial images in the dataset improves the accuracy of facial recognition. Further research is recommended to explore the effectiveness of the Haar Cascade Classifier method in facial recognition systems.