并行Adaboost-BP算法及其在海量图像分类中的应用  

The Parallel Adaboost-BP Algorithm and Application in Massive Image Classification

作　　者：曹建芳[1] 史昊[2] 赵青杉[1] 

机构地区：[1]忻州师范学院计算机科学与技术系,山西忻州034000 [2]太原科技大学计算机科学与技术学院,山西太原030024

出　　处：《新疆大学学报（自然科学版）》2017年第1期70-77,共8页Journal of Xinjiang University(Natural Science Edition)

基　　金：山西省自然科学基金(2013011017-2;2014011019-3);山西省高等学校重点教学改革研究项目(J2015099);忻州师范学院重点学科专项课题(XK201308)

摘　　要：为解决基于单节点架构的传统分类算法存在的时间效率差、分类准确率低等问题,提出了一种基于并行Adaboost-BP神经网络的海量图像分类方法.将BP神经网络作为弱分类器,由Adaboost算法组合15个BP神经网络的输出,构建了强分类器;充分利用了Hadoop平台下Map Reduce并行编程模型,提出了海量图像的自动分类模型,设计了并行Adaboost-BP神经网络算法的Map和Reduce任务.多组实验表明,相对于传统的AdaboostBP神经网络算法,提出的算法在Pascal VOC2007数据集和Caltech256数据集上的平均分类准确率分别提高了14.5%和26.0%,而且算法运行耗时少,系统加速比随集群节点个数增加而增加,在图像规模增加到20 000时,加速比几乎呈线性增长趋势.实验结果充分证明,提出的方法适合海量图像的自动分类和预测.In order to solve the problems of poor time efficiency and low classification accuracy existed in traditional classification algorithms based on single node architecture, a classification approach for massive images based on parallel Adaboost-BP neural network is proposed. Taken BP neural networks as weak classifiers, Adaboost algorithm combines the outputs of 15 BP neural networks and builds a strong classifier. The automatic classification model for massive image classification is put forward making full use of Map Reduce parallel programming model of Hadoop platform. The Map task and Reduce task of parallel Adaboost-BP neural network algorithm are designed. Multi group experiments show that the average classification accuracy of the proposed algorithm is increased by 14.5 and 26.0% respectively based on the VOC2007 Pascal dataset and Caltech256 dataset compared with traditional Adaboost-BP algorithm, and the time consuming of the proposed algorithm is less, the system speedup increases with the increase of the number of nodes in the cluster, especially, when the image scale increases to 20000, the speedup is almost linear. The experimental results fully demonstrate that the proposed approach is suitable for the automatic classification and prediction of massive images.

关 键 词：Adaboost-BP神经网络 图像分类 特征提取 MapReduce并行编程模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP317.4[自动化与计算机技术—计算机科学与技术]