Low-Power Scalable 3-D Face Frontalization Processor for CNN-Based Face Recognition in Mobile Devices

Abstract

A low-power scalable 3-D face frontalization processor is proposed for accurate face recognition in mobile devices. In spite of recent improvement in face recognition accuracy mainly from convolutional neural networks (CNNs), their performance is limited to face images with frontal view. For face recognition with human-level accuracy in real-life environment, in which most of the face images are captured from arbitrary angles, 3-D face frontalization is essential as a preprocessing stage for CNN-based face recognition algorithms. The proposed face frontalization processor shows scalability in two aspects: image resolution and accuracy. For low-power consumption and scalability, the processor proposes three features: 1) scalable processing element (PE) architecture with workload adaptation; 2) accuracy scalable regression weight quantization to reduce the external memory access (EMA) down to 81.3%; and 3) pipelined memory-level zero-skipping to further reduce the EMA by 98.4% without any latency overhead. From the proposed EMA reduction features, the EMA is reduced by 99.7% with little accuracy degradation in face frontalization results. The proposed face frontalization processor is implemented in 65-nm CMOS process, and it shows 4.73 frames/s) throughput. Moreover, power consumption of the implemented face frontalization processor is 0.53 mW, which is suitable for applications on mobile devices.