基于通道剪枝和量化的卷积神经网络压缩方法  

Convolution neural network comprehension method based on channel pruning and quantization

作　　者：徐晓 蒋志翔[1] 张杨[1] XU Xiao;JIANG Zhi-xiang;ZHANG Yang(Institute 706,Second Academy of China Aerospace Science and Industry Corporation,Beijing 100854,China)

机构地区：[1]中国航天科工集团第二研究院706所,北京100854

出　　处：《计算机工程与设计》2021年第10期2860-2866,共7页Computer Engineering and Design

摘　　要：为提升辅助维修技术在实际工程应用中的性能,解决嵌入式设备性能有限,难以实现实时物体识别任务这一问题,以提高轻量级卷积神经网络在嵌入式平台中的识别速度为目标,提出一种基于通道剪枝和量化的综合卷积神经网络压缩方法。以MobileNet V3模型进行实验,其结果表明,该卷积神经网络压缩方法有效压缩了网络结构,在识别精度损失可接受的情况下,实现了目标物体在嵌入式平台上的实时识别。To improve the performance of auxiliary maintenance technology in practical engineering application and solve the problems that the performance of embedded devices is limited and it is difficult to realize real-time object recognition task,a comprehensive convolutional neural network compression method based on channel pruning and quantization was proposed to improve the recognition speed of lightweight convolutional neural network in embedded platform.Experimental results on the MobileNet V3 model show that the convolutional neural network compression method effectively compresses the network structure,and realizes the real-time recognition of the target object on the embedded platform under the condition of acceptable loss of recognition accuracy.

关 键 词：卷积神经网络 物体识别 辅助维修 嵌入式应用 模型压缩 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]