Automatic Image Annotation via Combining Low-level Colour Feature with Features Learned from Convolutional Neural Networks

In this paper, a feature combination approach to annotate and retrieve images is proposed. In addition to using low-level colour features from original images, we extract features learned from convolutional neural networks (CNNs). We find these two sets are complementary to each other in conducting automatic image annotation (AIA). For both single-label CIFAR-10 and multi-label COREL-5K AIA tasks, the CNN-learned features perform slightly better than the low-level image features. Finally, when combining the two feature sets as inputs into the deep neural network-based AIA systems, we obtain the best performance in both cases.