基于卷积神经网络的多标签图像分类

Multi-label images classification based on convolutional neural network

针对BING算法对物体建模的不足,提出了multi-BING算法。该算法计算训练样本的CS-LBP特征,并对其进行聚类,对聚类后的数据建立BING模型。在物体检测过程中,融合了多个模型结果进行候选框判别,将多标签图像分类问题转化为多个单标签分类问题。以Fast R-CNN模型为基础,将采用本文物体检测方法得到的候选框作为模型输入。同时,采用LReLU函数作为Fast R-CNN模型的激活函数,从而在几乎不增加计算复杂度的情况下,提高模型的平均准确率(AP)。实验表明,本文方法优于BING算法和OBN算法。

In order to overcome the shortcoming of Binarized normed gradients(BING)algorithm in object modeling,a multi-BING algorithm is put forward.First,the Center-Symmetric Local Binary Pattern(CS-LBP)features of the training examples are calculated and clustering is performed.Then,different BING feature model is established based on different class of data.During the cause of object detection,all the results are emerged to find the candidates.The experiments show that the proposed method is significantly better than BING algorithm and Objectness(OBN)algorithm.In this,the problem of multilabel image classification is converted into the classification of many single-label images.Based on the Fast R-CNN model,the gained candidate box is taken as the input.At the same time,Leaky rectified linear unit(LReLU)function is considered as the activation function of the Fast R-CNN model.Thus,Average precision(AP)of the algorithm is promoted without more computing time cost and calculation overload.