BIDeepLab: An Improved Lightweight Multi-scale Feature Fusion Deeplab Algorithm for Facial Recognition on Mobile Devices
Jinming Li, Yutong Zhou
Abstract
This study presents BIDeepLab, a lightweight and improved multi-scale feature fusion algorithm based on DeepLab, specifically designed for facial recognition segmentation tasks on mobile devices. In response to the growing need for high-precision, low-latency face recognition in mobile applications—such as smart security, access control, and mobile identity verification—BIDeepLab introduces two key innovations. First, to address the challenge of multi-scale feature fusion during downsampling, we propose a Multi-Scale Attention (MSA) module that enables more efficient learning and integration of facial features at various scales. Second, inspired by the BiFPN architecture, we enhance the high-low feature fusion mechanism, allowing more accurate boundary and semantic information to be preserved during upsampling. These enhancements significantly improve segmentation quality while maintaining computational efficiency. Experiments were conducted on the Labeled Faces in the Wild (LFW) dataset, which includes over 13,000 real-world face images labeled with identities and detected using the Viola-Jones face detector. BIDeepLab achieved an mIoU of 90.2%, outperforming the original DeepLab in facial edge segmentation accuracy, while substantially reducing model parameters and computational cost. These results validate BIDeepLab as a practical and efficient framework for real-time facial segmentation on mobile and embedded systems.