Hardware Implementation of Convolutional Neural Network for Face Feature Extraction

As an important feed-forward neural network in the field of deep learning, convolutional neural network (CNN) has been widely used in image classification, face recognition, natural language processing and document analysis in recent years. CNN has a large amount of data and many multiply and accumulate (MAC) operations. With the diversity of application files, the channel sizes and kernel sizes of CNN are diverse, while the existing hardware platform mostly adopts the average optimization technology, which causes the waste of computing resources. In this paper, a special configurable convolution computing array is designed, which contains 15 convolution units, each PE contains 6x6 MAC operations, it can be configured to calculate three different kernel sizes of 5x5, 3x3 and 1x1. At the same time, pipeline structure is used to synchronize convolution and pooling operations, which reduces the storage of intermediate results. We design the special hardware structure to optimize DeepID network. Tested on Altera Cyclone V FPGA, the peak performance of each convolution layer at 50 MHz is 27 GOPS, and the average utilization of the MAC is 92%.