Block-wise mean squared error and dynamic margin for deep face recognition

State-of-the-art face recognition models, leveraging convolutional neural networks (CNNs) with margin-based Softmax approaches, deliver superior performance but incur significant computational overhead, posing challenges for deployment on resource-constrained edge devices. To address this, a novel knowledge distillation framework is proposed, specifically tailored for lightweight neural network architectures, integrating a Block-wise Mean Squared Error (MSE) loss and an enhanced ArcFace with dynamic margin adjustments. The Block-wise MSE partitions embedding features into weighted blocks based on variance analysis, prioritizing critical dimensions to optimize knowledge transfer for lightweight models. The dynamic margin in ArcFace adapts to sample difficulty, enhancing the learning of challenging samples (e.g., blurred or occluded images) to improve convergence and inter-class discriminability. Extensive evaluations across LFW, AgeDB-30, CFP-FP, CALFW, and CPLFW demonstrate superior accuracy and robustness compared to existing methods. Furthermore, the cosine similarity between teacher and student network outputs is validated as an effective difficulty indicator, providing a novel metric for assessing sample quality in distillation frameworks.