Design of deep convolutional neural network cascade for face detection

This paper proposes to build a high performance and reconfigurable face detection acceleration system, and uses the method of hardware and software collaboration to give full play to the advantages of ARM and FPGA. It accelerates the MTCNN cascaded deep convolutional neural network framework while ensuring recognition accuracy, and ultimately completes a face detection system based on ZYNQ. This design is mainly divided into two parts: software and hardware. In terms of software, facial detection adopts the MTCNN cascaded deep convolutional neural network framework. It uses a carefully designed three-layer cascaded deep convolutional network to predict facial positions and facial keypoint coordinates from rough to fine. The MTCNN model is trained on the PC using the Caffe framework, OpenCV visual library, and C++ language, using the WIDER FACE dataset for training. Afterwards, the system will be written and deployed on the ZYNQ 7020 SOC platform using standard C language (C99 standard) on the Xilinx SDK software. The hardware part combines the ADV7511 controller and video buffer to achieve real-time image display, displaying the images stored in the SD card. Using ARM Cortex-A9 hardcore processor and VDMA as AXI slave device, direct memory access of video data is achieved.