Optimized Face Detection and Alignment for Low-Cost and Low-Power IoT Systems

Face detection and alignment are challenging operations due to variations in image angles, background lighting conditions and intermediate blocking objects. Recent work has shown that these tasks can be improved through the use of a multi-task cascaded convolutional neural network (MTCNN) architecture. However, it is difficult to implement such an approach in a low-end edge AI system because of its high computational complexity. This paper presents the design of an MTCNN based on a low-cost and low-power processor/FPGA system that can be used in IoT applications. First, we analyze the computational requirements of the algorithm. Based on this analysis, we develop an optimized implementation to achieve real-time processing, taking advantage of the available hardware resources. In order to enhance the throughput and reduce the power consumption for AI edge devices, we store all intermediate results in on-chip block RAM. We achieve a frame rate of 15.2 frames per second, which meets the needs of security cameras that are widely used in IoT systems. Furthermore, our approach has a 2.67 times lower power consumption than for a previous MTCNN implementation.