An Accurate Weight Binarization Scheme for CNN Object Detectors with Two Scaling Factors

Abstract

Recently, convolutional neural network (CNN)-based object detectors such as You Only Look Once (YOLO) have been intensively studied for applications in robotics, drones, and autonomous driving. Although YOLO can run in real time by using a graphics processing unit, the YOLO hardware implementation has received a great deal of interest due to its power efficiency and the potential for massive chip production. However, extensive memory access and high computation complexity are widely known as bottlenecks in YOLO hardware implementation. A common and intuitive approach is to apply quantization, especially binarization, to object detectors. However, the existing binarization methods suffer from substantial degradation in detection performance. To address the problem, this study proposes an accurate weight binarization scheme using two scaling factors. Specifically, a new binary weight optimization problem is formulated, and an analytical solution is derived. Experimental results with well-known PASCAL Visual Object Classes show that the proposed method reduces the detection accuracy degradation by up to 32.18% while meeting the memory and computation requirements of state-of-the-art methods.