A Study On Deep Convolutional Feature Encoding

Recent studies have demonstrated the advantages of the representation learning ability of Convolutional Neural Networks(CNNs).Most existing methods adopt activations from the last fully connected layer as the image representation in the CNNs paradigm for visual classification tasks.This paper advocates exploiting appropriately deep convolutional feature from convolutional layer to constitute a powerful descriptor under an end-to-end learning framework.To this end,we propose two new architectures to utilize the convolutional feature before sent to classifier:The first architecture is a locality-aware coding layer conducted with the locality constraint,where the dictionary and the encoding representation are learned simultaneously.The locality-aware coding layer is readily amenable to training via the backpropagation as the locality-aware coding has an analytical solution.It is capable of capturing class-specific information which makes the learned convolutional features more robust.The resulting representation is particularly useful for texture classification.The second architecture is a kernel pooling layer which pools convolutional feature via a kernel matrix which makes samples lay in Riemannian manifold.It improves maxpooling and average pooling as it is a second-order pooling rather than first-order.It improves bilinear pooling as it employs non-linear kernel matrix.Kernel pooling enables network to capture order-less and fine-grained feature,thus is suitable for the texture classification and finegrained classification.Comprehensive experiments on several texture and fine-grained datasets show that our proposed architectures notably outperform the state-of-the-art methods.