Transfer Learning and Tuning of Deep Pre-trained Architecture for Face Recognition

Abstract

Automatic Image Identification is one of the interests of software developers with the application of machine and deep learning methods. With the incorporation of Transfer Learning and Tuning in pre-trained architecture, a substantial increase in the model’s performance is evident. This paper performs face recognition using an image identification and recognition approach. Feature extraction was performed using ResNet50 pre-trained architecture with Support Vector Machine as a classifier. Initial evaluation was made to generate a precision of 62.50%, recall of 65.55%, and f1-score of 63.99%. With this poor performance of ResNet50, the hyperparameters were tuned using transfer learning and tuning. After several times of manual experiments, a significant increase in precision is 93.75%, recall is 94.36%, and f1-score is 94.05%. Based on the remarkable yield of 35.25% for accuracy, 38.79% for recall, and an f1-score of 30.06%, it is advisable to apply the model for image identification and recognition