一种双LSTM结构的图像多标签分类方法

A bi-LSTM method for multi-label image classification

基于深度学习的图像多标签分类方法主要采用CNN-RNN方法进行按顺序标签预测。但是由于图像的多个标签没有特定的顺序,CNN-RNN方法难以确立有效的标签序列顺序,使得预测的精度不足。基于CNN-RNN方法构建了一种双LSTM结构,利用CNN从所给图像中提取出特征,构建两个LSTM同时解析图像特征,并采用不同的序列顺序进行预测,融合两个预测序列得到最终的预测标签。实验结果表明,文中提出的算法能够有效降低由单一的标签顺序带来的分类效果不足的影响,提高多标签分类的精度。

The CNN-RNN method is used in multi-label image classification based on deep learning to predict multiple labels with specific order. However,as the multiple labels of images do not have a specific order,it is difficult to establish an effective label sequence with the CNN-RNN method,which undermines the performance.Based on the CNN-RNN method,we built a bi-LSTM structure,using CNN extracting features from a given image,then constructed two LSTMs to analyze the feature at the same time,and predicted with different sequence order. Finally,we merged two predicted sequences to get the final prediction. The experimental results show that the proposed method can effectively reduce the effect of the insufficient classification caused by single sequence order and improve the performance of multi-label classification.