一种混合特征选择方法及应用研究  

An efficiently hybrid feature selection with its application

作　　者：赵学华[1] 刘学艳[1] 杨欣斌[1] 湛邵斌[1] 

机构地区：[1]深圳信息职业技术学院数字媒体学院,广东深圳518172

出　　处：《深圳信息职业技术学院学报》2016年第3期11-18,共8页Journal of Shenzhen Institute of Information Technology

基　　金：国家自然科学基金项目(61471133;61571444);广东省自然科学基金项目(2016A030310072);广东省高等学校优秀青年教师培养计划(YQ2013194);深圳市战略新兴产业发展专项资金(JCYJ20140418100633634)

摘　　要：针对目前基于机器视觉的棉花异性纤维在线检测的分类精度低和检测速度慢的问题,提出一种高效的棉花异性纤维混合特征选择方法。首先利用费舍尔评分滤波式特征选择方法过滤噪声特征,然后利用蚁群优化从已去噪的特征集中选取最优特征子集。提出的方法与费舍尔评分方法及基于蚁群优化的特征选择方法进行了对比分析,结果表明提出的方法选出的最优特征集仅包含12个特征,分类准确度达到93.45%,对一幅4000×500像素的彩色图像的在线检测时间仅为0.8116秒。所提方法能选择出具有较高分类精度、较小特征数量的优化特征子集,可以有效地改进棉花异性纤维检测的精度与效率,对提高现有基于机器视觉的棉花异性纤维在线检测设备性能具有重要的应用价值。For improving the classification accuracy and detection speed of foreign fibers in cotton based on machine vision, this paper proposed a hybrid feature selection algorithm for online detection of foreign fiber in cotton, which combines filter approach with wrapper approach together. First, Fisher Score is used to filter noisy features. Afterwards, the ACO uses the classifier accuracy as a fitness function to select the highly discriminating features. The proposed method is tested on foreign fiber dataset and make comparisons with Fisher Score and ACO. The experimental results shows that the proposed method is superior to Fisher Score approach and ACO. The selected subset only contains 12 features, and its classification accuracy get to 93.45% and the consuming time is only 0.8116s. The proposed method can efficiently select the optimal subsets with small size and high classification accuracy. The selected feature set can meet the requirement of online detection of cotton foreign fiber based on machine vision.

关 键 词：棉花异性纤维 混合特征选择 费舍尔评分 蚁群优化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]