基于FPGA的红外目标识别神经网络加速器设计

在红外目标识别领域，基于卷积神经网络的深度学习算法的识别精度已远远超过了传统模式识别算法,但神经网络的实现需要庞大的计算和存储，难以在无人机等嵌入式平台上进行部署。针对此问题，将通道级量化策略和梯度的近似优化训练引入到了低比特神经网络模型的建立中，并提出了一种可充分利用硬件计算资源的FPGA加速器，其整体平均性能为65.6GOPS。与其他相关工作的对比表明，低比特量化方法及其FPGA加速器实现，可以为嵌入式红外目标识别系统提供一种能效高、识别精度高的解决方案。

An Infrared Object Detection Neural Network Accelerator Based on FPGA

In the field of infrared object detection, the accuracy of deep learning algorithm based on convolutional neural network has exceeded that of traditional pattern recognition algorithms. However, the implementation of neural networks requires huge computing and storage demands, and it is difficult to deploy them on embedded platforms such as drones. To overcome the above problem, this paper introduces a channel-wise quantization strategy and the approximate optimization of the gradient into the training of low-bit neural network model. Besides, a hardware computing resources utilized FPGA accelerator is proposed to implement the model, whose overall performance is 65.6 GOPS. Comparisons with other related works show that the low-bit quantization method and the proposed FPGA accelerator implementation can provide a solution with high energy efficiency and high recognition accuracy for the embedded infrared object detection system.