Multi-label Image Classification Under Weakly Supervised Learning

Natural images are often multi-label images containing many types of objects.Accurate classification of multi-label images is not only one of the foundations of computer advanced visual understanding,but also has a wide range of applications in real life.Because multilabel images contain many kinds of objects and often have complex background information,it is a great challenge to classify them accurately.Strongly supervised learning is a good solution.However,a strongly supervised learning model requires expensive object-level or pixel-level labels as supervisory information.In order to reduce the cost of strongly supervised learning,this paper studies the multi-label image classification algorithm under weakly supervised learning.On the one hand,this paper improves the current mainstream image classification framework.In the current mainstream image classification framework,there is a serious feature competition between different types of objects.This kind of feature competition makes the number of features that can be obtained by a certain type in the process of classification severely affected by the number of samples of this type.Therefore,this paper proposes a framework with multi-path network to reduce the negative impact of feature competition among different types of objects.Experiments show that the proposed framework can alleviate feature competition and thus improve the performance of multi-label image classification algorithm.In addition,the proposed framework has good flexibility and can design the specific network structure according to different tasks.On the other hand,inspired by human visual attention mechanism,this paper designs a multilabel image classification model based on attention mechanism.Firstly,spatial attention mechanism is introduced into the multi-label image classification network to enable the network to better learn the spatial information of objects,which makes the network know “where to look”.Secondly,this paper proposes a multi-label image classification method based on channel attention,which takes into account the global information of channel features in convolutional networks,in order to highlight effective features,suppress noise features,and increase the correlation between features,so that the network can better understand “what to look”.Experiments show that the introduction of these two attention mechanisms in convolutional neural networks improves the performance of multi-label image classification model to a certain extent.The model with attention mechanism can accurately locate the target position.At the same time,compared with the backbone network,the attention model proposed in this paper only adds a few parameters,which has the advantages of lightweight and high efficiency.The research in this paper shows that the multi-path network framework can effectively deal with feature competition,and the introduction of attention mechanism can better extract object features in multi-label images.Both of them effectively improve the performance of multi-label image classification algorithm.In addition,the research in this paper provides a solution for exploring a lightweight and efficient multi-label image classification algorithm with an end-to-end manner.