摘要
针对林业业务图像的特点,提出了一种基于稠密尺度不变特征转换(Dense SIFT)特征的词袋(BoW)模型,并联合直方图正交核的支持向量机(SVM)对图像自动分类。首先采用Dense SIFT提取林业业务图像特征,然后使用BoW模型描述各业务图像,最后利用SVM进行分类识别。实验结果表明:采用Dense SIFT特征比SIFT特征训练时间和识别时间更短,并有更高的识别率,更适应实时性较高的场合;SVM采用多项式核函数(Poly),径向基核函数(RBF),多层感知器核函数(Sigmoid)以及直方图交叉核对3类林业业务图像分类时,直方图正交核取得的平均识别率最高;综合Dense SIFT在局部特征上的优势,加上BoW模型和直方图交叉核SVM分类器,平均识别率达到了86.7%,有较好的识别效果。
For characteristics of forestry images,an image classification method was put forward based on Dense SIFT and the BoW Model with support vector machine（SVM） using a histogram intersection kernel in order to improving to meet the need of the forest resources management.First,using the BoW Model,the Dense SIFT features of forestry images were extracted to describe the image.Then SVM was used for classification to identify the category of the images.Different kinds of kernel functions like Poly,RBF,Sigmoid,and the histogram intersection kernel were used to find the best recognition rate.Experimental results showed that using Dense SIFT had a shorter detection time（t=60.143 s） and a higher recognition rate（r=86.7%） than SIFT（t=95.567 s and r=83.3% respectively）,and it was suited for high real-time applications.Also the histogram intersection kernel had a higher average recognition（r=86.7%）.Combining Dense SIFT and the BoW Model with SVM and using the histogram intersection kernel,algorithms used with three kinds of forestry images had a better average recognition（r=86.7%）.
出处
《浙江农林大学学报》
CAS
CSCD
北大核心
2017年第5期791-797,共7页
Journal of Zhejiang A&F University
基金
浙江省自然科学基金资助项目(LY16C160007)
浙江农林大学科研发展基金人才启动项目(2013FR059)
关键词
森林计测学
林业业务图像
图像分类
特征提取
BoW模型
支持向量机
forest measurement
forestry images
image classification
feature extraction
BoW Model
support vector machine
