基于视频分析的视觉检测算法研究
Video Analysis-Based Research on Visual Detection Algorithms
王俊鹏 1蔡革英 1韩梅1
作者信息
摘要
本文针对运动烟支的动态监测、视频合成、缺陷分析等关键问题展开研究,并取得了一定理论和实用价值的成果.本文的工作包括:(1)视频合成的研究:一般工业相机采集的数据为单张的图像数据,并不能直接进行视频的保存.本研究采用FFmpeg开源视频处理库对采集到的图像进行视频合成,并利用视频的帧内压缩以及帧间压缩技术,成功实现海量图像的视频化动态存储.(2)基于卷积神经网络进行的深度学习研究:利用主流深度学习框架Pytorch对卷积神经网络的数据输入层、卷积计算层、激励层、池化层、全连接层等层进行研究并修改,在测试数据集上的实验结果表明了卷积神经网络在缺陷检测方面的优势.
Abstract
This article focuses on key issues such as dynamic monitoring,video synthesis,and defect analysis of moving cigarette sticks,and has achieved certain theoretical and practical value.The work of this article includes:1.Research on video synthesis.Data collected by typical industrial cameras are single image data and cannot be directly saved as video.In this study,the FFmpeg open-source video processing library was used to perform video synthesis on the captured images.Utilizing both intra-frame and inter-frame compression techniques,we successfully achieved the dynamic storage of massive images in video format.2.Research on deep learning based on convolutional neural networks.Research and modify the data input layer,convolutional computation layer,excitation layer,pooling layer,fully connected layer,and other layers of convolutional neural networks using the mainstream deep learning framework Pytorch.The experimental results on the test dataset demonstrate the advantages of convolutional neural networks for defect detection.
关键词
视频分析/视觉检测/烟支检测Key words
Video analysis/Visual detection/Cigarette stick detection引用本文复制引用
出版年
2024
NETL
NSTL
万方数据