Comparison of Tools to Locate Occluded Faces

• DOI: 10.22533/at.ed.31732723140810

• Palavras-chave: Location, Occluded Faces, Evaluation, Face Mask Detection, Artificial Vision.

• Keywords: Location, Occluded Faces, Evaluation, Face Mask Detection, Artificial Vision.

• Abstract:

  Derived from the pandemic, the World Health Organization (WHO) recommends using face masks as a way to contribute to reducing the spread of the virus that causes COVID-19; however, the use of the face mask causes many current facial biometric systems to be inefficient regarding the location and recognition of people when their faces are occluded with the face mask. This article presents a comparison between four of the main facial locators most used in people identification systems with pre-trained models, but applied to the detection of faces with occlusion caused by the use of a mask in uncontrolled environments. The tools that were evaluated are Dlib, MTCNN, DNN Facial Detector and MediaPipe. For evaluation, a sample of images was drawn from the public sets MaskedFace-Net, MFDD, RMFRD, SMFRD, and a small proprietary set with lighting changes, rotation, and various types of occlusion. After experimentation it was found that MediaPipe was better obtaining a precision of 97.5%, an accuracy of 86.7% and a 92.9% f1 score in this work.