Learning Label Correlation For Multi-label Image Recognition

The multi-label image classification aims to accurately classify the categories of different objects in the image,which are widely used in real scenes such as scene recognition and attribute classification.The core research issue of this task is to exploit the correlation among labels.Learning the correlation among labels is a long-standing problem in multi-label image classification task.The label correlation is the key to solve the multi-label classification but it is so abstract that hard to model.The rapid development of deep learning technology has accelerated the research process of multi-label image classification,and the powerful fitting ability of deep neural network provides a new label correlation learning method.For the learning problem of label correlation,we designed new multi-label recognition algorithm to map images and labels to the same space.In this space,metric learning mechanism is used to cluster the related features and labels.We also designed Constrain ranking Loss to supervise the learning process of the latent space.However,most solutions try learning image label dependencies to improve multi-label classification performance.But they have ignored two more realistic problems: object scale inconsistent and label tail(category imbalance).These two problems will impact bad influence on the classification model.To tackle these two problems and learning the label correlations,we proposed a solution named Feature Attention Network(FAN).Based on self-Attention mechanism,FAN pay more ‘attention' on more important features and learn the correlations among convolutional features to indirectly learning the label dependencies.Following our proposed solutions,we achieve best multi-label image classification performance on MSCOCO 2014 and VOC 2007 dataset compared with other great solutions.