Image classification method is mainly to classify the extracted image features using the classifier. Hence,the extracted image features and the classifier used directly affect the classification results. Generally,the artificial feature extraction models are used in im-age feature extraction. However,the effective artificial feature extraction models are difficult for the images with complex contents. Moreover,with the increase of the size of the training set,it requires a huge amount of training time to get better classification accuracy results. To address these issues,the hybrid neural network classifier is proposed in this paper and it is combined with deep belief net-works for image classification. The hybrid neural network classifier is constituted by the evolution function element layer and the neu-ron layer. In the hybrid neural network classifier,the evolution function element layer is as the input layer and the neuron layer is as the output layer for the classification results. Deep belief networks are graphical models which automatically extracts the deep hierarchical representation of the input data. The new image classification method proposed in this paper consists of 2 steps. First,the deep belief network which is formed by a stack of restricted boltzmann machines is used to extract the feature vectors of the images,and secondly, the hybrid neural network classifier is used to classify these feature vectors. Experimental verifications are conducted on MNIST dataset and UCI dataset. Experimental results indicate that compared with the combination of the stacked restricted boltzmann machines and softmax classifier,and the combination of the stacked restricted boltzmann machines and ES-based softmax classifier and the combina-tion of the stacked restricted boltzmann machines and support vector machine,our approach can get the higher classification accuracy in less time and have superior anti-overfitting ability.%图像分类方法主要是使用分类器对提取的图像特征进行分类.因此,提取的图像特征和使用的分类器直接影响分类结果.图像特征提取一般是人为设定特征提取模式,然而,对于内容复杂的图像难以人为设定有效的特征模式.此外,随着训练集规模的增加,分类器想要获得更好的分类精度需要大量的训练时间.为了解决这些问题,提出混合神经网络分类器,并将该分类器和深度置信网络结合设计了新的图像分类方法.混合神经网络分类器由演化函数模块层和神经元层组成,演化函数模块层作为输入层,神经元层作为分类结果的输出层.深度置信网络是一种用于自动提取输入数据深层特征的网络模型.本文中提出的新的图像分类方法分为2个步骤,首先,堆叠受限玻尔兹曼机构成的深度置信网络用于提取图像的特征向量,其次,使用混合神经网络分类器对提取的特征向量进行分类.采用MNIST数据集和UCI数据集对提出的方法进行实验验证.实验结果表明,与堆叠受限玻尔兹曼机和softmax分类器的组合,堆叠受限玻尔兹曼机和基于演化策略的softmax分类器的组合以及堆叠受限玻尔兹曼机和支持向量机的组合相比,提出的分类方法可以在更短的时间内获得比较高的分类精度并且具有更好的抗过拟合能力.
展开▼
