基于标签相关性的卷积神经网络多标签分类算法

Convolutional Neural Network Multi-label Classification Algorithm Based on Label Correlation

针对卷积神经网络对于多标签分类中未考虑标签相关性的问题,提出了一种考虑标签相关性的卷积神经网络分类方法。在Alex Net网络结构的基础上进行了两处改进:首先,通过共享卷积层的通用特征,降低神经网络的训练时间;其次,根据贝叶斯模型计算待预测标签的后验概率,通过在神经网络的目标函数中引入后验概率与预测概率的近似度作为损失项,提升多标签分类的准确率。实验结果表明,所提算法在Cars车辆数据集上取得了更好的分类效果。

Since the Convolutional Neural Network algorithm ignores the correlation between labes,a convolution neural network classification method by exploiting label correlation is proposed in this paper.There are two improvements based on the AlexNet network structure.Firstly,By sharing common features of the convolutional layer,training time of the neural network is reduced.Secondly,according to the Bayesian model,the posterior probability of the label to be predicted is calculated.