多分类器融合的光学遥感图像目标识别算法-《计算机技术与发展》

文章信息/Info

Title:: Optical Remote Sensing Image Object Recognition Based on Multiple Classifications Fusion

Author(s):: JI Xiao-fei; SHI Yu-chen; School of Automation,Shenyang Aerospace University,Shenyang 110136,China

Keywords:: optical remote sensing image; decision fusion; Hog feature; Zernike feature; support vector machine; BP neural network; random forest

摘要:: 光学遥感图像的多目标检测与识别一直是图像处理与分析的热点研究问题。基于单一特征单一分类器的多目标光学遥感图像分类识别算法存在识别准确率不高的问题。对此,充分利用特征与识别方法之间的适应性,提出了一种多特征多分类器融合的光学遥感图像多目标识别算法。首先对光学遥感图像的分类目标提取 2 种具有平移、缩放不变性的特征表示:Hog 特征和 Zernike 特征;其次分别用 3 种适应性较好的分类器(BP 神经网络、支持向量机(SVM)、随机森林(RF))进行分类;最后在决策级分别融合两种特征、三种分类器的概率输出,给出最终的分类结果。实验结果表明,该算法较大程度地提高了光学遥感图像多目标识别的准确性,对飞机、舰船、油罐、汽车四类多目标的识别取得了 95.37% 的正确识别率。

Abstract:: The multi-target detection and recognition of optical remote sensing images has always been a hot topic in image processing and analysis. The classification and recognition algorithm of multi-target optical remote sensing image based on single feature single classifier has a low recognition accuracy. For this,an optical remote sensing image multi-target recognition algorithm based on multi-feature and multi-classifier fusion is proposed by making full use of the adaptability between features and recognition methods. Firstly,two kinds of features with translation and scaling invariance are extracted from the classification target of optical remote sensing image: Hog feature and Zernike feature. Secondly,three kinds of better classifiers (BP neural network,support vector machine (SVM),random forest (RF)) are used for classification. Finally the final recognition results by using decision level probability fusion are given. The experiment indicates that this algorithm improves the accuracy of multi-target recognition of optical remote sensing images to a large extent,and achieves a correct recognition rate of 95. 37% for aircraft,ship,oil tank and automobile.