Selective Feature Exclusion for Masked Face Recognition in ResNet-34 Deep Learning Architecture

Our research presents a novel approach to enhance the recognition of individuals wearing masks, which can be particularly challenging due to obscured facial features. By incorporating our method into the ResNet-34 deep learning framework, we aim to overcome difficulties in detecting masked faces in specific frames. Typically, a mask complicates facial recognition, as it merges all features, including the mask itself, into a singular facial representation. We used a selective feature extraction strategy to address this, focusing on specific face landmarks. We tested 5, 25, and 31-point landmarks within the ResNet-34 model, adapting our method to improve masked face detection. Additionally, we utilized an adaptive technique and refined our feature selection process, supporting the enhancement of detection rates by focusing on smaller subsets of facial landmarks when a face is not initially recognized. Further accuracy improvements were achieved by employing a Support Vector Machine (SVM) in the model's final layers, providing a refined approach to feature differentiation without relying on principal component analysis (PCA) due to the mask's influence. We also explored variations in ensemble model structure, finding an optimal configuration for face recognition involving a shallower cascade regressor. Our best results were obtained with settings of 31 landmark points, a 15- depth cascade, and a tree depth of 4, successfully identifying 300 out of 359 masked faces.