A face detection and recognition method built on the improved MobileFaceNet

Face recognition has increasingly become the predominant biometric recognition technology for identity verification, propelled by advancements in deep learning technology. This study introduces a lightweight face detection and recognition method optimised for mobile devices with limited computational resources using an improved MobileFaceNet framework. Initially, the approach refines the network structure, elevating face detection efficiency through median filtering and a minimal bounding box constraint strategy grounded in the multitask convolutional neural network (MTCNN). Subsequently, to address the challenges of multi-pose in real-world scenarios of face detection, the method employs Affine Transformation for facial angle rotation and centre point adjustment, thus achieving accurate pose correction in facial images. The study presents a lightweight face recognition network model based on MobileFaceNet in its final phase. It improves the model by optimising the loss function and learning rate and reducing convolutional layers by integrating depthwise separable convolution. In addition, regarding the privacy security of face recognition information, it proposes a face information encryption scheme built on a fully homomorphic encryption algorithm. Experiments on prevalent face databases demonstrate that this model is better in recognition accuracy and network performance.