B-Face: 0.2 MW CNN-Based Face Recognition Processor with Face Alignment for Mobile User Identification

An ultra-low-power face recognition processor with binary weight convolutional neural network core and face alignment accelerator is proposed for high accuracy even with head pose variations. Binary convolution core with exhaustive input reuse and interleaved output memory access is proposed to minimize power consumption, resulting in 13.3 TOPS/W power efficiency. In addition, face alignment core with zero-aware pipelining is proposed to minimize external memory access. As a result, the face recognition system with maximum 48 frames-per-second throughput and 0.2 mW minimum power consumption is realized.

Introduction

Recently, face recognition (FR) is widely adopted to mobile devices as a measure for user identification, enabling user-spe-cific services. The convolutional neural network (CNN) shows high accuracy in FR, so that a dedicated CNN FR processor was developed to reduce power consumption in battery-limited environment [1]. However, variance in head pose causes accuracy degradation for stand-alone CNN up to 10% [2], as shown in Fig. 1. Face alignment which converts the input face into its frontal face can enhance the recognition accuracy, but its software realization on mobile application processors consume >1W power, greatly diminishing battery run-time. The necessity of the face alignment hardware was first raised in [3], but only simulated results were given. Therefore, a low-power FR processor with face aligning capability is necessary for high accuracy mobile FR system.