基于自适应深度置信网络的图像分类方法

Image classification algorithm based on adaptive deep belief network

针对传统的深度置信网络在数据的特征表达过程中收敛速度较慢、训练时间较长的问题,提出一种非监督学习算法,即自适应深度置信网络,将其应用于图像分类任务中。采用一种自适应步长大小方法,解决训练时对适合的学习率的选择困难,加速训练的收敛性。在MNIST手写数据集上进行测试验证并与多个分类器的性能进行对比分析,实验结果表明,该算法具有更快的收敛速度和良好泛化能力,图像的分类效果得到有效提升。

The feature representation of data is often a computationally expensive process that involves slow convergence rate and requires a considerable amount of training time for tradition deep belief network.To mitigate this problem,an unsupervised learning algorithm,namely adaptive deep belief network,was proposed and applied to image classification tasks.An adaptive step size technique was used,which solved the difficulty of choosing an adequate learning rate and accelerated the training convergence.The testing verification was conducted on the MNIST dataset and the contrastive analysis was carried out compared with several classifiers＇ performance.The results show the proposed algorithm has higher convergence speed and better generalization ability,and the results of image classification are also effectively improved.