摘要
现有的图像跟踪方法直接对图像像素级特征展开建模,未考虑图像内部深层视觉特征信息,导致其难以准确描述轮廓信息,致使跟踪效果较差。为解决上述问题,基于深度学习设计了新的双阈值图像局部分块视觉跟踪方法。以深度卷积神经网络为基础构建PigNet网络,检测图像分辨率与位置信息。然后通过PLSA算法估计分割区域类别,筛选出候选区域,并获得双阈值空间信息。继而利用EM算法在贝叶斯算法基础上更新高斯模型参数,获得最优模型参数和最大后验概率,从而实现视觉跟踪。实验结果表明,上述方法在双阈值图像局部分块视觉跟踪方面误差更小,证明新方法的实用性能较强。
Generally, the traditional image tracking methods ignore deep visual feature information inside images, resulting in false contour information and poor tracking effect. In this regard, this paper designed a new dual-threshold image local block visual tracking method based on deep learning. First, based on the deep convolution neural network, the PigNet network was constructed to detect the image resolution and position information. Second, according to the PLSA algorithm, the region categories were segmented to screen out candidate regions, obtaining the dual-threshold spatial information, and then, based on EM and Bayesian algorithms, the parameters of the Gaussian model were updated to obtain the optimal model parameters and maximum posterior probability. Finally, visual tracking was achieved. The results show that this method has a minor error and outstanding practicability in the visual tracking of the local block of double-threshold images.
作者
韩开旭
袁淑芳
HAN Kai-xu;YUAN Shu-fang(College of Electronics and Information Engineering,Beibu Gulf University,Qinzhou Guangxi 535011,China;College of Sciences,Beibu Gulf University,Qinzhou 5 Guangxi 35011,China)
出处
《计算机仿真》
北大核心
2021年第5期172-175,共4页
Computer Simulation
基金
北部湾大学引进高层次人才科研启动项目(2018KYQD35)
广西高校中青年教师科研基础能力提升项目(2020KY10019)
广西高校中青年教师科研基础能力提升项目(2021KY0434)。
关键词
双阈值图像
深度卷积神经网络
深度学习
Double threshold image
Deep convolution neural network
Deep learning
作者简介
韩开旭(1984-),男(汉族),黑龙江大庆人,博士,讲师,主要研究领域为数字图像处理、自然语言处理;通讯作者:袁淑芳(1988-),女(汉族),河北衡水人,硕士,助理研究员,主要研究领域为数字图像处理。
