面向FPGA部署的CNN-SVM算法研究与实现

卷积神经网络-支持向量机(CNN-SVM)混合算法结合了CNN特征提取能力和SVM分类性能,在计算复杂度和解决小样本问题上具有一定优势,目前已在故障诊断、医学图像处理等领域得到了一定应用,同时,由于其计算复杂度较低,也引起了边缘计算领域的关注。针对边缘计算场景中对算法性能和功耗的要求,提出了一种面向FPGA平台的CNN-SVM算法优化与实现方法。首先,结合FPGA的架构特点,对CNN-SVM算法结构进行了硬件适应性优化,包括模型压缩和分类器核函数的选取。其次,采用了软硬件协同和高层次综合(HLS)设计方法,完成了CNN-SVM算法加速器的设计与实现。实验结果表明,在ZCU102上,加速器的FPS(frames per second)达到了18.33 K,计算速度为1.474 GMAC/s,相对于CPU平台四核Cortex-A57和Ryzen7 3700x分别实现了23.57和4.92倍加速,相对于Jetson Nano GPU和GTX750平台能耗比分别达到了33.24和50.27。

Research and implementation of CNN-SVM algorithm based on FPGA

CNN-SVM hybrid algorithm combines the feature extraction ability of CNN and the classification performance of SVM, it has certain advantages in computational complexity and can solve small sample problem. It has been applied in fault diagnosis, medical image processing and other fields, at the same time, it gets attention in the field of edge computing due to its low computational complexity. Aiming at the requirements of algorithm performance and power consumption in edge computing scenarios, an optimization and implementation method of CNN-SVM algorithm for FPGA platform is proposed. First, combined with the architecture characteristics of FPGA, the hardware adaptability optimization of CNN-SVM algorithm structure is carried out, including the model compression and the selection of kernel function of classifier. Secondly, the design and implementation of CNN-SVM algorithmic accelerator is completed by using software and hardware cooperation and high level synthesis ( HLS) design method. The experimental results show that on ZCU102, the frames per second(FPS) of accelerator reaches 18. 33 K, the computing speed is 1. 474 GMAC/ s. Compared with the CPU platform, quad core Cortex-A57 and Ryzen7 3700x achieve 23. 57 and 4. 92 times acceleration respectively, compared with Jetson Nano GPU and GTX750 platform, the energy consumption ratio is 33. 24 and 50. 27 respectively.