NoonGil Lens+: Second Level Face Recognition from Detected Objects to Decrease Computation and Performance Trade-off

Abstract

Artificial intelligence has developed in various fields. The development became more significant after Neural Networks(NN) began to gain popularity. Convolutional Neural Networks(CNNs) are good at solving problems such as classification and object detection. However, the CNNs model tends to function to solve a specific problem. In the case of both object detection and face recognition it is difficult to make a single model that works well. NoonGil Lens+ is expected to be an approach that can solve both problems at once. As well as being a solution, it is also hoped that this approach can reduce the trade-off of accuracy and execution speed. The approach we propose can be called as Noongil Lens+, a system that connects YOLOv3 and FaceNet. It is inspired from a korean series called ‘STARTUP’. The author only develops the FaceNet model and the proposed system in this paper (NoonGil Lens+). Region Selection, a machine learning-based greedy approach was proposed to determine snapshots to fed into FaceNet for facial identity classification. FaceNet is trained on the CelebA dataset which has gone through the preprocessing process and is validated using the LFW dataset. NoonGil Lens+ was validated using 70 images of 7 celebrities, characters, and athletes. In general, the research was carried out successfully. NoonGil Lens+ using Region Selection has an accuracy of up to 75.2%. The Region Selection execution speed is also faster compared to Cascade Faces.