Fully automated multi-label image annotation by convolutional neural network and adaptive thresholding
Hoa M. Le, Thi-Oanh Nguyen, Dung Ngo-Tien · 2016
This paper presents a fully automated and flexible ConvNet-based classifier for multi-label image annotation. The classifier alleviates hierarchical representation of image from a convolutional neural network, and adaptive thresholding technique on the ranked list of label scores. The method can annotate images with arbitrary number of labels that the classifier finds fit, as opposed to common methods which only assign a fixed number of those. Experiments show state-of-the-art on classification accuracy and competitive annotation performance across 2 intrinsically different data-sets, Corel5K and MSRCv2. Although the proposed method shows some limitation in learning label semantics, empirical study indicates that it was due to the established drawback of univariate loss function, which the classifier optimised, in multi-label classification. It, therefore, opens for number of directions to improve the performance, while still retains the merits of the proposed method.