Hybrid Fixed-Point/Binary Deep Neural Network Design Methodology for Low-Power Object Detection

Suffering from both high computational complexity and high memory bandwidth is the major challenge in realizing the deep neural network in low-power for real-time applications. Binarizing the feature maps as well as the filter coefficients in deep neural network is an efficient way to reduce the high power consumption in deep learning object detection, however, it greatly scarifies the detection accuracy when reducing the bit-width of a 32-bit word to a binary bit in a floating-point deep neural network. This paper proposes a hybrid fixed point/binary deep neural network design methodology for object detection to achieve low-power consumption by taking advantage of both the fixed-point and binary deep neural networks, which allocates enough bit-width to design the hardware datapath in different layers of deep neural network. The proposed methodology combines dynamic fixed-point quantization and binarization techniques together to extremely compress the object detection model to result in a compact hybrid fixed-point/binary detection neural network, which achieves lower bandwidth and lower computational complexity. An automation tool based on the proposed methodology is also developed to train a hybrid deep neural network under a specified quality loss range. Taking MobileNet-SSD as an example, using the proposed methodology, the resulted model achieves 91% model size reduction and 75.8% memory bandwidth reduction at the cost of less than 1% mAP quality degradation. The proposed design methodology for hybrid fixed-point/binary deep neural networks achieves a good balance on detection accuracy, model size compression ratio and feature map reduction for low-power deep learning object detection applications.