《DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition》
一,主旨
这篇论文主要讨论在训练样本不足的情况下,从深度卷积网络提取出来的特征是否能移植到其他的目标识别任务中,释放了DeCAF开源深度学习平台
二,结构
1,对特征有两方面分析
1)通篇使用的是在ILSVRC-2012 validation set上,从alexnet中提取出来的特征,与其他方法提取出来的特征如(a) LLC , (b) GIST, and features derived from our CNN: (c) DeCAF1, the first pooling layer, and (d) DeCAF6相比较,发现DeCAF6有更好的聚类效果。
2)在时间消耗上,CNN要比传统方法提取特征耗时长,且卷积层,全连接层消耗时间多,所以一些稀疏方法的引入对减少时间的消耗非常有用
2,四个实验,分别从basic object recognition, domain adaptation, fine-grained recognition, and scene recognition.方面阐述全连接层特征的泛化能力。
1)Object recognition
on the Caltech-101 dataset
训练时dropout,测试时,所有activations都乘0.5
实验表明,训练特征可被移植,加上线性SVM都能取得很好的效果
2)Domain adaptation
数据集Office dataset
例从Dslr上训练得到的特征检测Webac,取得很好的效果,说明特征的可移植性
3)Subcategory recognition
on the task of subcategory recognition ,the Caltech-UCSD birds dataset
indicating that such features, although not specifically designed to model subcategorylevel differences, captures such information well
4)Scene recognition
on the SUN-397 large-scale scene recognition dataset
DeCAF is learned on ILSVRC, an object recognition database, we are applying it to a task for which it was not designed
三,结论
Our visual results demonstrate the generality and semantic knowledge implicit in these features, showing that the features tend to cluster images into interesting semantic categories on which the network was never explicitly trained.
四,提炼
ALEXNET,输入图片大小的改变,裁取方式的改变;特征可以是网络中提取的特征与其他特征的结合
***个人理解,错误的地方请指出,谢谢!