DEEP LEARNING METHODS BASED ON MobileNetV2 MODEL FOR LOW-LIGHT FACE RECOGNITION

Abstract

In this research paper, the problem of reduction of the recognition accuracy for face image captured in low-light environments in modern recognition methods is considered. Reviewed recognition methods are based on deep learning methods for extracting object of interest (a face), digital transformation of an image and different convolutional neural networks models (VGG-Face, ArcFace, FaceNet, DeepFace, SphereFace, CosFace, OpenFace) for feature extraction and classification. Two methods are proposed for facial recognition, based on the MobileNetV2 deep learning model architecture with the integration of additional blocks for feature extraction enhancements. One of the proposed methods includes prior classification of images depending on the lighting conditions of an image, based on the usage of a CNN with accuracy of 99.33%, after which a correspondingly-trained MobileNetV2 architecture-based model is used for recognition itself. Proposed methods were tested on the UTKFace dataset, as well as generated DistortionFace and NormalFace datasets, for recognition of face images by features such as age, gender and ethnicity. Each dataset consists of 23 708 images in total, which are then divided into parts of 16 595 images for training, 4 741 for validation and 2 371 for testing. Proposed methods for facial recognition based on MobileNetV2 showed mean absolute error of 6.38 for age, as well as accuracy of 88.09% for gender and accuracy of 72.93% for ethnicity. Proposed method for face recognition based on a lighting classification CNN model and MobileNetV2 model showed mean absolute error of 6.74 for age, as well as accuracy of 87.14% for gender and accuracy of 70.53% for ethnicity. Proposed methods demonstrate robustness in low-light face images recognition and potential for further research and improvements.