基于包级空间多示例稀疏表示的图像分类算法

发布时间：2018-12-25 19:38

【摘要】：基于多示例学习框架的图像分类算法以其特有的多义性对象表示能力在图像分类中表现出较好的分类效果。但传统的包级空间多示例学习算法在特征选择过程中存在忽略小目标概念区域且包含大量冗余信息的问题,造成部分训练包信息损失,影响分类性能。为此,基于多示例学习与稀疏编码理论提出1种改进的多示例图像分类算法。该算法首先根据同类样本示例聚为一簇的特性,应用聚类算法构造每类图像的视觉词汇,并利用负包中所有示例都为负的特性,对视觉词汇进行约束,消除冗余信息;依据训练样本示例与视觉词汇的相似度,获得每类训练样本的包特征向量。然后,基于稀疏编码理论,对训练包中的包特征向量进行稀疏编码,获得每1类训练样本的字典矩阵。最后,对待分类样本特征进行稀疏线性组合,预测待分类样本的类别标签。通过对COREL数据集图像进行测试,结果表明,与其他多示例学习算法相比,文中提出的方法能较好地解决图像分类问题,具有较高的分类精度。
[Abstract]:The image classification algorithm based on multi-example learning framework has better classification effect in image classification because of its unique multi-object representation ability. However, the traditional multi-example learning algorithm in packet-level space has the problem of neglecting the concept area of small target and containing a lot of redundant information in the process of feature selection, which results in the loss of part of the training packet information and affects the classification performance. Therefore, an improved multi-example image classification algorithm based on multi-example learning and sparse coding theory is proposed. The algorithm firstly constructs the visual vocabulary of each kind of image by using the clustering algorithm according to the characteristics of the same kind of sample examples clustered into a cluster, and uses the negative feature of all the examples in the negative packet to constrain the visual vocabulary and eliminate redundant information. According to the similarity between training samples and visual vocabulary, the packet feature vectors of each kind of training samples are obtained. Then, based on the sparse coding theory, the feature vector of the training packet is sparse coded, and the dictionary matrix of each training sample is obtained. Finally, a sparse linear combination is used to predict the class labels of the classified samples. By testing the images of COREL dataset, the results show that the proposed method can solve the problem of image classification better and has higher classification accuracy than other multi-example learning algorithms.
【作者单位】：西北工业大学自动化学院;
【基金】：航天科技创新基金(CASC201104) 航空科学基金(2012ZC53043)资助
【分类号】：TP391.41

【相似文献】