基于多特征融合的深度置信网络图像分类算法

Image Classification Algorithm for Deep Belief Network Based on Multi-feature Fusion

针对现有单一特征描述符及浅层结构分类算法分类正确率较低的问题,基于底层图像特征提出一种针对自然界图像特点的深度置信网络(DBN)图像分类算法。提取样本图像中的颜色、纹理和形状特征,构成多特征融合的权重矩阵,并对特征矩阵进行归一化处理,利用构建的4层DBN分类器进行训练和分类。采用Corel图库,通过训练权重进行测试,结果表明,该算法的平均分类正确率达到85.1%,高于使用单一特征的分类算法和其他主流分类算法。

Taking the single feature and the major classification algorithms into consideration, an image classification algorithm based on fusion of multi-feature for Deep Belief Network（DBN） is proposed to classify the nature images. The features about color, texture, shape are extracted and the characteristic weight matrix is formed. Then the characteristic matrix is normalized. The samples are trained and classified using the DBN with four levels which is constructed. The proposed method has been evaluated on the Corel dataset by train weight, and the result shows that the average classification accuracy is 85.1% by the proposed algorithm,which is higher then single feature algorithm and other mainstream algorithms.