Comparative Analysis of MTCNN and Haar Cascades for Face Detection in Images with Variation in Yaw Poses and Facial Occlusions

As computer vision and machine learning advance, face detection has become a major focus. Face recognition has several methods and models. Every implementation starts with face detection. Haar Cascades and Multi-task Cascaded Convolutional Networks (MTCNN) are compared for facial pose variation robustness. This research will examine how well these two models detect faces in yaw postures from -90 to +90 degrees. Many studies have contrasted these two models, but the yaw poses of faces were not addressed due to the scarcity of datasets with systematic degrees of face orientation. Thus, the UPM face dataset, created at the UPM embedded systems lab using developed equipment to produce high-resolution photographs and a systematic range of face orientations from -90 to 90 degrees, was used to evaluate the range of degrees these two models can reach. UPM includes 100 students with different yaw angles and occlusions (masks, glasses, or both). The results reveal that MTCNN is the best for detecting faces with yaw poses only, masks, glasses, and both at all degrees (-90 to +90) with 100%, 99.9%, 96.4%, and 80% accuracy. Instead, Haar cascades were 92.5%, 67.3%, 80.4%, and 76.3% accurate.