Mid Sweden University

This study explores the development of a robust Single Sample FaceRecognition (SSFR) system leveraging deep learning techniques. The focus is on overcoming the challenges associated with limited data, where only one image per person is available for training. Various convolutional neural network (CNN) models, including VGGFace,VGG-16 ResNet-50, and MobileNet, have been evaluated for their accuracy and scalability. Advanced preprocessing techniques, such as multi-scale, multicrop,

flipping, and Principal Component Analysis (PCA), have been used to increase model performance. The final model, using VGGFace for feature extraction and K-Nearest Neighbors (KNN) for classification, achieved an accuracy of 88.13% on the Labeled Faces in the Wild (LFW) dataset.

Scalability tests indicated that the model’s training time increased linearly with the number of training samples, demonstrating practical applicability for large-scale implementations. The study shows the importance

of applying suitable preprocessing and modular approaches in developing scalable and accurate SSFR systems.