Reducing and Stretching Deep Convolutional Activation Features for Accurate Image Classification

In order to extract effective representations of data using deep learning models, deep convolutional activation feature (DeCAF) is usually considered. However, since the deep models for learning DeCAF are generally pre-trained, the dimensionality of DeCAF is simply fixed to a constant number (e.g., 4096D). In this case, one may ask whether DeCAF is good enough for image classification and whether we can further improve its performance? In this paper, to answer these two challenging 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 and increase its dimensionality by stretching the weight matrix between successive layers. To improve the performance of RS-DeCAF, we also present a modified version of RS-DeCAF by applying the fine-tuning operation. Extensive experiments on several image classification tasks show that RS-DeCAF not only improves DeCAF but also outperforms previous “stretching” approaches. More importantly, from the results, we find that RS-DeCAF can generally achieve the highest classification accuracy when its dimensionality is two to four times of that of DeCAF.