Providing a hybrid method for face detection, gender recognition, facial landmarks localization and pose estimation using deep learning to improve accuracy

Subject Areas : journal of Artificial Intelligence in Electrical Engineering

peyman  jabraelzadeh ¹ (Department of Electrical Engineering Ahar branch, Islamic Azad university, Ahar, Iran)
Asghar  charmin ² (Department of Electrical Engineering,Ahar Branch, Islamic Azad University, Ahar, Iran)
Mohsen  Ebadpore ³ (Department of Electrical Engineering,Ahar Branch, Islamic Azad University, Ahar, Iran)

Received: 2022-01-14 Accepted : 2022-01-27 Published : 2020-03-01

Keywords: face detection, Convolutional network, Yolo, Pose estimation, Gender recognition,

Abstract :

In general, identifying and locating faces in images or videos is considered as the first step in face recognition. It is quite clear that an accurate detection algorithm can significantly benefit system performance and vice versa. Therefore, face recognition is one of the key steps in the application of face recognition systems. In deep learning algorithms are able to learn high-level features, which have been highly regarded by researchers for use in the field of machine vision, as well as in a variety of fields such as image classification and human gesture estimation, which are the key activities for image perception. In this paper, we present a hybrid method called Hyper-Yolo-face to identify faces, facial landmarks localization, pose estimation and recognize the gender of a given image using deep convolutional neural networks, the Yolo algorithm, and local binary patterns. The proposed network architecture is based on the AlexNet model and the integration of the binary pattern operator and Yolov3, which results in increasing performance and accuracy. Yolo changes the architecture of face recognition systems and looks at the problem of recognition as a regression problem which goes directly from the pixels of the image to the coordinates of the box and the probability of the classes. Experiments on the AFLW and FDDB datasets indicated that the proposed model performs significantly better than other algorithms and methods and improves detection accuracy.