Image classification based on deep local feature coding

In this paper, we propose an improved locally aggregated descriptor (VLAD) algorithm coded on deep local features for image classification. Firstly, convolutional neural network (CNN) is adopted to extract the dense local features of images. Secondly, a subset of feature, chosen by the criterion of normal distribution, is selected for high quality codebook generation. Finally, the local features are assigned to multi-neighbor visual words instead of the nearest one with different weights, simultaneously, the statistical distribution information about local features is taken into account during VLAD coding process. Extensive experiments on public available datasets demonstrate the promising performance of the proposed method against state-of-the-art methods.