<![CDATA[Multi-face detection presents a significant challenge in computer vision, especially in environments with limited hardware resources. This study compares the performance of three multi-face detection methods—Haarcascade, Dlib (HOG and CNN), and RetinaFace—using a subset of the WIDER FACE dataset in a CPU-only environment without GPU acceleration. The experiment was conducted in two stages using a total of 300 images from the WIDER FACE dataset, which reflect real-world variations such as pose, scale, illumination, expression, and occlusion. Performance evaluation was carried out using precision, recall, F1-score, accuracy, and processing time as metrics. The results show that RetinaFace consistently outperforms the other methods, achieving superior metrics in Recall (0.92), F1-score (0.93), and Accuracy (0.88) on Subset A, and leading across all metrics on Subset B. While Dlib-CNN demonstrates high detection performance, it suffers from very slow processing time. In contrast, Haarcascade delivers the fastest processing speed but performs poorly in terms of evaluation metrics. The experiments also reveal that RetinaFace is the most consistent and reliable method based on standard deviation values of precision (0.01), recall (0.11), F1-score (0.07), and accuracy (0.11). Overall, this study contributes valuable insights for selecting efficient face detection methods under constrained resource conditions.]]>
Copyrights © 2025
