Completion of Missing Labels for Multi-Label Annotation by a Unified Graph Laplacian Regularization

• DOI: 10.1587/transinf.2019edp7318
• Publication Date: 2020-09-30T22:31:59Z
• Authors: Jonathan MOJOO; Yu ZHAO; Muthu Subash KAVITHA; Junichi MIYAO; Takio KURITA
• Keywords: Regularization

The task of image annotation is becoming enormously important for efficient retrieval from the web and other large databases. However, huge semantic information complex dependency labels on an make challenging. Hence determining similarity between multiple useful to understand any incomplete label assignment retrieval. This work proposes a novel method solve problem multi-label by unifying two different types Laplacian regularization terms in deep convolutional neural network (CNN) robust performance. unified model implemented address missing efficiently generating contextual both internally externally through their similarities, which main contribution this study. Specifically, we generate matrices using Hayashi's quantification method-type III word2vec method. generated methods are then combined as term, used new objective function CNN. Regularization term study able problem, enabling more effectively trained network. Experimental results public benchmark datasets reveal that proposed with CNN produces significantly better than baseline without state-of-the-art predicting labels.