Occluded face recognition algorithm based on MFFPN with lightweight network

Face recognition is a biometric technology used to identify individuals by extracting their facial features. The current face recognition research based on deep learning in limited scenarios has made significant progress and is extensively applied in portable terminals like smartphones and laptops. However, in complex situations, such as posture changes accompanied by shadows or occlusions in facial images, some facial features are missing, resulting in poor performance of traditional face recognition algorithms and low recognition accuracy. To address the aforementioned problems, the approach in this study uses the lightweight MobileFaceN et (MFN) as the basic network, and then formulates a novel network structure (MFFPN) through fusion with the Feature Pyramid Network (FPN) structure, in order to combine low-level and high-level features more efficiently to acquire more comprehensive facial features. Furthermore, the integration of FPN causes a Complexity of the network, leading to overfitting of the network as the network size and computation increase. To solve this issue, the Dropout regularization technique is implemented to randomly deactivate a proportion of neurons in the network, allowing the network to reduce its size and computational requirements while avoiding overfitting. Subsequently, to enhance the network's generalization capabilities and stability, the PReLU activation function in the original network is replaced with the Mish activation function. The experimental results demonstrate that the final MFFPN in occluded facial recognition improves performance to a certain extent when comparing with MFN and other conventional networks.