Abstract:
Face recognition is a fundamental problem in numerous real-world applications, and it can be tackled using classification deep learning models. This kind of solution requ...Show MoreMetadata
Abstract:
Face recognition is a fundamental problem in numerous real-world applications, and it can be tackled using classification deep learning models. This kind of solution requires extensive datasets to represent the target subjects, where maintaining their privacy is essential. Further, there are challenging scenarios where the face of the subject can be partially occluded, e.g. a subject wearing a mask, making it difficult the recognition process. In this paper, extensive experiments are presented for deep learning-based face recognition, considering different dataset settings. More precisely, cropping operations are proposed to partially represent the face focusing on the eyes and nose region for model training, minimizing privacy issues, storage space, and allowing the trained model to process partially occluded faces. Two publicly available datasets named VGG-Face and DigiFace-1M are adapted for evaluation, and five convolutional neural network models are used for comparison, including ResNet, VGG-16, AlexNet, DenseNet-169, and EfficientNetV2-M. The results suggest that, although this cropping operation impacts the accuracy of the model when compared to the processing of full faces, it can be viable solution to minimize the pointed issues while preserving acceptable performance.
Date of Conference: 09-11 July 2024
Date Added to IEEE Xplore: 19 August 2024
ISBN Information: