Adaptive Upsampling and Optimal Parameterization of Support Vector Machine for Enhanced Face Recognition in Resnet-based Deep Learning Framework

Face recognition has advanced significantly due to deep learning, primarily via libraries like Dlib. This paper proposes adaptive upsampling in the Residual Network (ResNet)-34 deep learning architecture to improve face detection in low-resolution images. The Support Vector Machine (SVM) is integrated into the fully connected layers classification tasks to improve accuracy. Principal Component Analysis (PCA) reduces feature dimensions, reducing training time and increasing detection rate. ResNet-34 in Dlib is trained using 128 face feature vectors and Euclidean distance on the Labeled Faces in the Wild (LFW) dataset. Reducing PCA dimensions from 100 to 50 or SVM cost parameters from 100 to 50 has no noticeable impact on accuracy. Still, it gives more Frame Per Second (FPS) in real-time recognition because the parameters are fewer.