An Optimal Flow for Face Recognition: Analysis/Effects of Face Detection, Alignment and Cropping Techniques

Facial recognition systems are commonly used to improve security and user experience. Facial biometrics is a technology that recognizes and verifies individuals based on unique anatomical features. Since 2012, deep learning methods have accelerated the development of face recognition systems, making them faster, more reliable, and successful. While the literature on face recognition systems has become mature, new methods and algorithms are continuously being introduced to improve the robustness of these systems. However, it is important to note that each system has its own advantages and disadvantages depending on its intended use, making it difficult to derive a universally successful system. This study aims to analyze the effect of pre-processing methods such as face detection, image cropping, image resizing, image normalization on the success of face recognition algorithms using modified modules for face recognition systems in the industry. The study identifies the most efficient method and the combination of modules that will increase success. The study found that using the proposed method in the pre-processing stage improved the success of all flows. The recommended modules for the best flow are RetinaFace for the face detection, with the proposed approach for alignment and scaling, and the ArcFace model for feature extraction. The simulation studies demonstrated that it achieved superior results on the LFW dataset compared to other methods, with an accuracy of 97.7