基于低比特量化神经网络的红外目标识别方法及其FPGA实现

A Low-Bit Quantization Neural Network Method for Infrared Object Detection and Its FPGA Implementation

红外目标识别系统成为航空航天、无人驾驶等军事和民用领域中一项至关重要的技术。红外目标识别算法是红外目标检测识别系统中的核心之一。传统红外目标识别技术往往依赖人为的特征选择,无法对复杂困难的红外目标实现高效、准确的识别。本文提出了训练中反量化与通道级量化相结合的量化策略,有效减小量化误差对网络模型性能的影响。实验结果表明:本文提出的低比特量化算法在红外数据集上有着优异的表现。在硬件部署方面,本文提出了更加高效的卷积计算单元,提高了硬件资源的利用率,同时也达到了更高的峰值性能。最终,在PYNQ-Z2嵌入式现场可编程门阵列(FPGA)上进行验证,系统在150 MHz的时钟频率下达到了90.6 GOP/s的峰值吞吐率,其功耗为2.5 W。

The infrared object detection system has become a vital technology in military and civilian fields such as aerospace and unmanned driving. The infrared object detection algorithm is one of the cores in the infrared object detection and recognition system. In this paper, a quantization strategy combining the learnable quantization in training and channel-level quantization is proposed, which effectively reduces the effects of quantization errors on the performance of the network model. The test results show that the proposed low-bit quantization algorithm has an excellent performance on the self-built infrared dataset and the public visible light dataset. Besides, an efficient convolution calculation unit is proposed for the hardware deployment, which improves the utilization of the hardware resources and the peak performance. Finally, verification tests are carried out on the PYNQ-Z2 embedded field programmable gate array(FPGA). The results show that the system reaches a peak throughput of 90.6 GOP/s at a clock frequency of 150 MHz, and its power consumption is 2.5 W.