基于深度学习混合模型迁移学习的图像分类  

Image Classification Based on Mixed Deep Learning Model Transfer Learning

作　　者：石祥滨[1,2,3] 房雪键 张德园[1] 郭忠强[3] 

机构地区：[1]沈阳航空航天大学计算机学院,沈阳110136 [2]沈阳航空航天大学辽宁通用航空重点实验室,沈阳110136 [3]辽宁大学信息学院,沈阳110036

出　　处：《系统仿真学报》2016年第1期167-173,182,共8页Journal of System Simulation

基　　金：国家自然科学基金(61170185);航空科学基金(2013ZC54011);辽宁省博士启动基金(20121034);辽宁省教育厅资助项目(L2014070)

摘　　要：为提高深度模型迁移学习的特征识别力,提出一种基于受限玻尔兹曼机与卷积神经网络混合模型迁移学习的图像分类方法。该方法融合了2种模型特征的学习能力,提取图像的结构性高阶统计特征进行主题分类。该方法在迁移预训练的卷积神经网络模型到小目标集时,使用受限玻尔兹曼机代替卷积神经网络模型中的全连接层,在目标集上重新训练受限玻尔兹曼机层和Softmax层,并使用BP算法进行参数调整。加入的受限玻尔兹曼机层不仅全连接所有特征maps,还从最大对数似然的角度学习目标集特有的统计特征,消除了数据集间内容差异对迁移学习特征识别力的影响。在Pascal VOC2007和Caltech101数据集上的实验结果表明,该方法具有较高的分类准确率。In order to obtain high discrimination image representations in limited amount of datasets, the method based on mixed deep transfer learning model was proposed. When trained CNNs transferred to the target datasets, fully-connected layers were replaced by RBM layers. The method retrained the RBM layers and Softmax classifier, then fine-tuned the mixed model with backpropagation algorithm. The RBM layers not only fully connected whole feature maps, but also learned the target datasets＇ statistical features in the view of the biggest logarithmic likelihood, to eliminate the effects caused by the content differences between datasets. The experimental results show that the method has improved the accuracy of image classification, outperforming other methods on Pascal VOC2007 and Caltech101 datasets.

关 键 词：图像分类 卷积神经网络 受限玻尔兹曼机 迁移学习 Softmax 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]