基于区域划分的DBN分类及其在舰船目标识别中的应用

A Classification Method of Region-based DBN and Its Application in Ship Recognition

针对舰船目标识别时效性需求,论文提出了一种基于局部区域划分的DBN舰船目标识别方法。该方法首先提出了一种基于像素灰度级区域划分的局部空间灰度信息特征表征方法,有效衡量了当前像素与其搜索窗口内像素的邻域结构差异性,理论分析表明,该方法时间复杂度较低;其次构建了深度信念网络模型,通过反向传播网络向前传播计算得到样本所在的类别索引,实现了目标的准确分类;最后仿真实验表明,该算法的分类性能优于其他算法:K近邻、支持向量机、分层判别回归以及传统深度信念网络,满足舰船目标识别处理的高效性要求。

Aiming at the requirement of effectively recognizing ships using optical remote sensing data,a ship recognition method of deep belief network(DBN)based on region partition is proposed. Firstly,according to region partition based pixel gray levels,the expression of feature extraction based on local spatial gray information of the pixel is built,which can efficiently measure similarity between the current pixel and its neighborhood structure in a search window,and is able to produce stable and accurate ship recognition feature. In addition,theoretical analysis shows that the expression of feature extraction has reduced time complexity. Secondly,a classification method based on(DBN)model is presented,of which is combined by three restricted Boltzmann machine(RBM)and back-propagation network(BPN),it can learn features automatically and efficiently. At last,experimental results demonstrate that the proposed method can obtain higher recognition rate than k-nearest-neighbor(KNN),support vector machine(SVM),hierarchical discriminant regression(HDR)and traditional DBN model,it satisfies the time efficiency demand of ship recognition.