基于网络方法的形状图像特征选择  

Shape Image Feature Selection Based on Network Methods

作　　者：杨波 李国华 李金海 YANG Bo;LI Guohua;LI Jinhai(Data Science Research Center,Kunming University of Science and Technology,Kunming 650500,China;Faculty of Science,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学数据科学研究中心,云南昆明650500 [2]昆明理工大学理学院,云南昆明650500

出　　处：《昆明理工大学学报（自然科学版）》2023年第6期30-38,共9页Journal of Kunming University of Science and Technology(Natural Science)

基　　金：国家自然科学基金项目(11971211).

摘　　要：在形状图像的表示和识别领域,复杂网络方法不依赖于像素点的具体位置,在对具有旋转、缩放和拉伸等特征的形状进行分类时具有显著优势.本文首先提取形状图像的轮廓点作为网络的节点,然后依据轮廓点间的欧氏距离建立连边,最后选择网络的归一化平均度、最大度、熵和能量作为图像的特征进行分类.为了提高分类准确性,通常同时考虑不同阈值对应的多个轮廓点网络.本文着重研究了网络特性随阈值的变化规律以及轮廓点和阈值数量对分类准确性的影响.得到了一些有趣的结论:(1)内部有孔洞的形状图像其分类准确率低于内部无孔洞的形状图像;(2)不同类别形状之间的相似度越大,轮廓点数量的鲁棒性越差;(3)随机森林(RF)对不同形状数据集的分类均具有良好的稳定性.In the field of shape image representation and recognition,complex network methods,which do not rely on the specific positions of pixels,have significant advantages in classifying shapes with features such as rotation,scaling,and stretching.This paper first extracts the contour points of shape images as nodes of the network,establishes edges based on the Euclidean distance between the contour points,and finally selects the normalized average degree,maximum degree,entropy,and energy of the network as features for image classification.To improve classification accuracy,multiple networks corresponding to different thresholds of contour points are usually considered simultaneously.This paper focuses on studying the variation of network characteristics with thresholds and the influence of the number of contour points and thresholds on classification accuracy.Some interesting conclusions are drawn:(1)Shape images with internal holes have lower classification accuracy than those without holes;(2)The robustness of the number of contour points decreases as the similarity between different categories of shapes increases;(3)Random Forest(RF)exhibits good stability in the classification of different shape datasets.

关 键 词：轮廓点网络 滑动窗口方法 形状分类 特征选择 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]