关于视觉特征与CapsNet的图像大数据分类研究

Research on Image Big Data Classification Based on Visual Features and Capsnet

随着信息技术的快速发展,每天都有数以万计的图像产生,如何从中挖掘重要的图像信息是当前研究的热点问题之一,对此提出了一种关于视觉特征与CapsNet的图像大数据分类方法。为了解决大量图像数据的计算复杂度过高,以及灰度颜色直方图中没有对图像位置的问题,将图像灰度进行压缩,并采用共生矩阵和分形维数对视觉特征进行提取。采用胶囊网络中神经元的输出来表达图像中所包含的各种属性信息,为了更新胶囊网络的耦合系数,通过动态路由算法表示胶囊与子胶囊间的关系,在训练和测试中对动态路由不断进行计算得出胶囊网络的输出。把图像大数据分类算法部署到云计算节点上,采用批量更新的数据模型,将图像的训练集划分为众多数据块进行Map并行训练,利用训练样本向前、后传播得出权值梯度,并采用Reduce计算出所有训练样本权值梯度的平均值,同时对样本权值进行更新。实验结果表明,提出的方法可以有效地防止图像过拟合现象发生,图像分类的准确率和效率均有明显地提高,在图像大数据分类方面表现出显著的性能优势。

With the rapid development of information technology, tens of thousands of images are produced every day. How to mine important image information from it is one of the hot issues in current research. In this paper, a method of image big data classification based on visual features and information is proposed. In order to solve the high computational complexity of a large number of image data and the problem that there is no image position in the gray color histogram, the gray level of the image was compressed, and the co-occurrence matrix and fractal dimension were used to extract the visual features. The output of neurons in the capsule network was used to express various attribute information contained in the image. In order to update the coupling coefficient of the capsule network, the dynamic routing algorithm was used to express the relationship between the capsule and the sub capsule, and the dynamic routing was continuously calculated in the training and testing to obtain the output of the capsule network. The image big data classification algorithm was deployed to the cloud computing node, and the batch update data model was adopted. The training set of the image was divided into many data blocks for parallel training, and the weight gradient was obtained by forward and backward propagation of the training samples. The average value of all training samples weight gradient was calculated, and the sample weights were updated. Experimental results show that the proposed method can effectively prevent the occurrence of image overfitting phenomenon, improve the accuracy andefficiency of image classification, and show significant performance advantages in image big data classification.