Image multi-label learning algorithm based on label correlation

• Published in: 2021 IEEE International Conference on Consumer Electronics and Computer Engineering (ICCECE) pp. 606 - 609
• Main Authors: Huang, Mengyue, Zhao, Ping
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.01.2021
• Subjects: Classification algorithms; Computational modeling; Correlation; Data models; label correlation; multi-label classification; Probability distribution; Residual network-tree; Semantics; Training data
• DOI: 10.1109/ICCECE51280.2021.9342484

Aiming at the problem of image multi-label classification, the number of sample label categories is large, the output space of corresponding multi-label classification increases exponentially, and the training data is lacking. This paper proposes an image multi-label learning algorithm based on the label correlation residual network-tree model. The algorithm is based on the residual network-tree model for each label category in the sample corresponding to a branch, and independently trains a classifier; the semantic correlation between the labels in the sample is used to select training data for the classifier, and avoid the interference of the missing labels in the sample to the classifier, while at the same time train with the residual network-tree model. The experiment was conducted on the large-scale multi-label data set: Pascal VOC 2007 images. And the results showed that the algorithm proposed in the article was superior to mainstream multi-label classification algorithms in the classification effect of experimental data sets.