Face detection and recognition in an unconstrained environment

• Main Author: Chin, Ariana Sue Rei
• Format: Final Year Project / Dissertation / Thesis
• Published: 2022
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://eprints.utar.edu.my/4955/1/3E_1900599_FYP_Report_%2D_CHIN_SUE_REI_ARIANA.pdf; http://eprints.utar.edu.my/4955/

Face detection and recognition in an unconstrained environment is a challenging subject despite a plethora of solutions and the earnest efforts of numerous academics. Face refers to the front part of the head in humans from the forehead to the chin which includes the mouth, nose, cheeks, and eyes. Face detection refers to the ‘ability’ to identify faces whereas face recognition refers to the automated ‘technique’ used to confirm or identify a person based on physiological traits. Unconstrained environment refers to the lack of control over external variables such as illumination, position, occlusion, and distance from the camera. Face detection and recognition is a non-intrusive, discreet method of authenticating personal identification within enforcement agencies and business settings. While each face detection and recognition procedures is effective for the specific variant under investigation, performance declines quickly when other variances are present. In this project, five face datasets; (i)Labeled Faces in the Wild (LFW), (ii)Adience, (iii)Unconstrained Facial Images (UFI), (iv)Open Images V6 and (v)Unconstrained Face Detection Dataset (UFDD) with training set of 1500 images and test set of 300 images taken in an unconstrained environment (includes rain, snow, haze, blur, illumination and lens impediments) are manually annotated and trained on face detection models YOLOv4 and YOLOv5. It was found that YOLOv5 perform similarly with YOLOv4 using the mAP metrics. However, the training time of 6000 iterations for YOLOv5 is significantly lesser than YOLOv4. Additionally, YOLOv5 produces a much smaller weight file of 14 MB compared to YOLOv4. Open Images dataset performed the best with YOLOv5 model (86.1% mAP) for 300 test images taken in an unconstrained environment. The larger dataset LFW with greater number of labelled individual (1573 individuals) was able to achieve satisfactory results when verified with known or unknown faces in the database using the Siamese Neural Networks. The Siamese Neural Network can be further improved by increasing the number of verification images for each individual to reduce the likelihood of the model predicting a false positive. Finally, a prototype is developed using the face detection model, YOLOv5, using the weights obtained from Open Images dataset, and the face recognition Siamese Neural Network model, with the weights from LFW dataset.