Is DeCAF Good Enough for Accurate Image Classification?

In recent years, deep learning has attracted much interest for addressing complex AI tasks. However, most of the deep learning models need to be trained for a long time in order to obtain good results. To overcome this problem, the deep convolutional activation feature DeCAF was proposed, which is directly extracted from the activation of a well trained deep convolutional neural network. Nevertheless, the dimensionality of DeCAF is simply fixed to a constant number. In this case, one may ask whether DeCAF is good enough for image classification applications and whether we can further improve its performance? To answer these two questions, we propose a new model called RS-DeCAF based on "reducing" and "stretching" the dimensionality of DeCAF. In the implementation of RS-DeCAF, we reduce the dimensionality of DeCAF using dimensionality reduction methods, such as PCA, and meanwhile increase the dimensionality by stretching the weight matrix between successive layers. RS-DeCAF is aimed to discover the effective representations of data for classification tasks. As there is no back propagation is needed for network training, RS-DeCAF is very efficient and can be easily applied to large scale problems. Extensive experiments on image classification show that RS-DeCAF not only slightly improves DeCAF, but dramatically outperforms previous "stretching" and other state-of-the-art approaches. Hence, RS-DeCAF can be considered as an effective substitute for previous DeCAF and "stretching" approaches.