FARVIS: A Facial Attribute Recognizer Based on Vision Sensor Using Super-Efficient Network Supporting Service Robots

Service robots have developed into a significant tool in advertising industry and facial attribute recognition is crucial because it allows robots to understand human preferences, intentions, and emotions. Robot machines typically utilize a low-cost device such as a central processing unit (CPU) as the primary control unit for efficient implementation. Therefore, efficient facial attribute recognition that does not demand extensive memory and computational resources is required. This work proposes a super-efficient facial attribute recognizer (SEFAR) designed specially to support service robots. SEFAR introduces a super-efficient network consisting of a bottleneck multi-perspective convolution (BMPC) reinforced by an aggregated large kernel attention (ALKA) module. SEFAR, which has low computation and few parameters, achieves highly competitive performance compared to other networks on the CelebA and labeled faces in the wild attributes (LFWA) as benchmark datasets. By integrating the proposed SEFAR with face detection as the initial process and a camera as a vision-based sensor, this work proposes facial attribute recognition based on vision sensor (FARVIS). It operates in real-time at a maximum speed of 21.63, 51.90, and 86.58 frames per second (FPS) using face input images of a ${224} \times {224}$ , ${112} \times {112}$ , and ${56} \times {56}$ , respectively. FARVIS achieves this performance on a video graphics array (VGA) camera resolution using an Intel Core i7-9750H CPU without demanding extensive memory or computational resources. The demo video is presented at https://bit.ly/FARVIS-IEEESensors-AP