摘要
针对自然图像分类算法的精度低以及网络训练耗时过长的实际问题,提出了一种结合卷积自动编码器(Convolutional Auto-Encoders,CAE)的改进堆叠自动编码器(Stacked Auto-encoders,SAE)网络。研究了CAE学习局部特征的能力,并将其作为整个SAE网络的第一层。在提取初步特征的同时降低输入的维度,解决了网络参数过多,训练过程慢的问题。同时对改进的SAE网络进行微调,缩减训练时间,并提取更有利于分类的图像高层特征。实验结果表明,改进SAE网络对于自然图像的分类具有更好的普适性,可以有效地提高分类准确度,并加快网络训练速度。
In order to solve the natural image problems of low classification accuracy and the long time of the network training,an improved stacked auto-encoders( SAE) with the combination of convolutional auto-encoders( CAE) is proposed in this paper. By analyzing the ability of CAE for learning local features,CAE is used as the first layer of the whole SAE network,which extracts the local features and reduces the dimension of input to solve the problem of excessive parameters and the slow training process.At the same time,the SAE network is fine-tuned to reduce the training time and extract more meaningful high-level image features. The experimental results of the algorithm show that the improved SAE network is generally applied for the variety of natural images, can improve the classification performance effectively for natural images and the training speed of the network.
出处
《信息技术》
2016年第8期1-4,8,共5页
Information Technology
基金
国家自然科学基金(41306089)
江苏省产学研前瞻性研究项目(BY2014041)
常州市科技支撑项目(CE20145038)
关键词
图像分类
改进SAE网络
卷积自动编码器
微调
最大池化
image classification
improved SAE network
convolutional auto-encoders
fine-tuning
max pooling
