摘要
随着机器视觉技术发展,对识别速度、准确率和项目开发周期等方面都提出了更高的要求。人工智能无疑是较好的解决方式,而以往从底层编程搭建深度学习框架在技术和项目进度上都很难满足要求。为了满足图像识别要求,选择专用的图像处理服务器进行训练与识别,并对其主要部件进行选型。对SSD模型结构进行了分析,根据各层次结构计算了一次训练与前向运算过程中所需计算的参数及内存要求。通过开源的深度学习框架TensorFLow、SSD识别模型,在Python环境下设计图像格式转换、图像识别和定位程序。并在VOT2016标准数据集中进行测试。测试结果显示,在速度和识别成功率上都符合预期要求。
With the development of machine vision technology,higher requirements are put forward for recognition speed,accuracy and project development cycle.Artificial intelligence is undoubtedly a better solution,but in the past,building a deep learning framework from the bottom programming is difficult to meet the technical and project progress requirements.In order to meet the requirements of image recognition,a special image processing server is selected for training and recognition,and its main components are selected.The structure of SSD model is analyzed,and the parameters and memory requirements for training and forward operation are calculated according to the hierarchical structure.Using SSD recognition model of machine learning framework TensorFLow,the program of image format conversion,image recognition and localization is designed with Python.It is tested on VOT 2016 standard data set;the test results show that the expected requirements in terms of speed and recognition success rate are achieved.
作者
姜华
孙勇
Jiang Hua;Sun Yong(Zhejiang Economic Information Center,Hangzhou,Zhejiang 310000,China;Zhejiang Vocational and Technical College of Communications)
出处
《计算机时代》
2019年第6期71-75,共5页
Computer Era
基金
浙江省2018年度重点研发计划项目(2018C01111)
