Image classification with an RGB-channel nonsubsampled contourlet transform and a convolutional neural network

Lingling Fang, Hanyu Zhang, Jiaxin Zhou, XiangHai Wang

Abstract: In this paper, an efficient image classification method is proposed that is based on the nonsubsampled contourlet transform (NSCT) of RGB-channel images and the convolutional neural network (CNN). First, the NSCT-based coefficients of natural RGB-channel images are extracted, which are capable of capturing the statistical properties of each channel. In addition, the proposed feature descriptor is equipped with the mean–max-pooling strategy according to the characteristics of the correlated coefficients. Then, the CNN is concatenated to exaggerate the discriminative parts of the primary features. With these advantages, the proposed RGB-channel NSCT–CNN should, in general, improve the corresponding CNN-based image classification methods. Using the Food-101 and SUN Datasets, the proposed method achieves state-of-the-art classification results that are also significant for object detection. In addition, the proposed method can achieve better or comparable accuracy compared to other related methods based on these two datasets.