Abstract:

Bag Of Words (BOW) is a classical approach of image description, and the method of constructing the characteristic dictionary in this model is very important. A category constrained low-rank optimization characteristic dictionary training approach named LRC-DT was proposed for the characteristic dictionary construction. Through the low-rank optimization, the rank of the coefficient matrix constructed by same category images was minimized. Then the classification information was introduced into the characteristic dictionary learning to improve the identifiability of characteristic dictionary for image description. Some experiments were conducted on two standard image databases including Caltech-101 and Caltech-256, and the characteristic dictionary of SPM (Spatial Pyramid Matching), ScSPM (Sparse codes SPM), LLC (Locality-constrained Linear Coding) and LSPM (Linear SPM) were replaced by constrained low-rank optimization characteristic dictionary. The experimental results show that the proposed method can consistently offer better performance than not employing the category constrained low-rank optimization, its classification accuracy is improved with the increase of the training sample number.

摘要：

字典模型(BOW)是一种经典的图像描述方法，模型中特征字典的构造方法至关重要。针对特征字典构造问题，提出了一种类别约束下的低秩优化特征字典构造方法LRC-DT，通过低秩优化的方法使训练出来的特征字典在描述同类图像时表示系数矩阵的秩相对较低，从而将类别信息引入到字典学习中，提高字典对图像描述的可分辨性。在标准公测库Caltech-101和Caltech-256上的实验结果表明：将SPM、稀疏编码下的SPM(ScSPM)、局部线性编码(LLC)和线性核函数的SPM(LSPM)编码方法中的特征字典替换为加入低秩约束(LRC)的特征字典后，随着训练样本数目增多，字典模型的分类准确率与未引入低秩约束的方法相比有所提高。