基于多尺度主线方向特征的深度信念网络图像分类算法.docVIP

基于多尺度主线方向特征的DBN分类方法 高 强，李 倩 （华北电力大学电气与电子工程学院，河北 保定 071003） 摘要：针对深度信念网络（DBN）结构自身未考虑到二维图像空间结构信息、分类准确识别率不高等问题，提出了一种基于多尺度主线方向特征的DBN网络图像分类新方法（MSMD-DBN方法）。该方法首先提取多尺度的线方向特征图和能量图，再通过二值化、细化能量图得到主线方向特征图，然后在可视层输入端加入多尺度主线方向信息特征图，并利用深度信念网络进行图像分类识别。旨在通过增加输入信息的维度，来达到提升图像分类性能的目的。在CIFAR-10和MNIST两个数据库上对不同样本的图像进行分类实验，结果表明，与采用传统DBN网络和DBN的改进算法相比，本文算法的分类性能取得了显著的提高。 关键词：深度信念网络；多尺度主线方向特征；图像分类；正确识别率；分类性能 中图分类号:TN919.85 文献标志码:A An Classification Method Based on Multi-scale Main Direction Feature of Deep Belief Network GAO Qiang，LI Qian Institute of Electrical and Electronic Engineering，North China Electric Power University，Hebei Baoding 071003，China） Abstract: The structure of deep belief network doesn’t take into account the spatial information, and has poor recognition performance. In order to solve these problems, the feature extraction method of multi-scale main direction was introduced, and a new algorithm named MSMD-DBN was proposed. At first, multi-scale main direction feature maps and energy diagrams were extracted, and main line direction feature maps obtained by binarization and thinning. Then, the multi-scale main direction features were added into the visible layer. Finally, these features were inputted into the model of deep belief network, and the method was applied into the image classification. This paper focuses on increasing the dimension of the input information by adding main line direction features, so as to improve the classified performance. The recognition results on CIFAR-10 and MNIST databases with different samples show that, compared with the traditional deep belief network method and one of the improved methods of deep belief network, the proposed method achieves good recognition results. Key words: deep belief network; multi-scale main direction feature; image classification; correct recognition rate; classif