Face Recognition System with Occlusion Based on Attention Mechanism Improvement of the Vision Mamba Model^*

With the outbreak of COVID-19, the widespread use of face masks has become a major challenge for face recognition systems. The issue mainly arises from feature loss due to occlusion, especially in regions like the eyes and forehead, with common occlusions including masks, sunglasses, and winter hats. These significantly reduce the accuracy of face recognition. While visual Transformers have shown strong performance in recognition tasks, their quadratic computational complexity affects training efficiency. To address this, the Transformer-based Mamba network has been proposed, which offers an O(n) computational complexity and a hardware selection mechanism, significantly improving training efficiency. This paper presents an improved method based on the Vision Mamba network, which incorporates channel attention mechanisms to enhance the model's ability to handle occluded face recognition tasks and improve both training effectiveness and recognition accuracy.