Multi label image classification learning with weak labels

Abstract

A new multi-label similarity semantic learning(ML-SSL) model was proposed to solve the problem of label missing in existing multi-label image classification methods, It can produce better classification results by effectively recovering missing label information in training data. The model considers the characteristics of label structure and instance features, and recovers the missing label information in training data by using the label correlation within images and the similarity between images. After label recovery, the new training set is used to train the classification model, and the model is used to predict the test set. The experimental results show that the model has better performance improvement in image classification tasks under weak labeling.