基于平均区域划分的Laplacian稀疏编码的图像分类

IMAGE CLASSIFICATION BASED ON AVERAGE REGION PARTITIONING AND LAPLACIAN SPARSE CODING

针对稀疏编码方法中编码过程不稳定和金字塔匹配的划分方法无法使得融合后的特征很稀疏这两个问题,提出基于平均区域划分的Laplacian稀疏编码LSCARD(Laplacian sparse coding based on average region division)的图像分类方法。首先,对原始图像进行局部不变特征转化(SIFT)特征提取;然后,在稀疏编码方法中加入Laplacian正则化对局部特征进行编码,使相似的特征具有相似的码字;再利用平均区域划分以及最大值融合将编码后的特征向量进行融合;最后,采用多类SVM分类器对图像进行分类。在几个标准图像数据集上的实验结果表明,LSCARD算法具有更高的分类精度。

For the sparse coding method, the coding process is unstable and the pyramid matching method can not make the fusion feature very sparse, an image classification method based on Laplacian sparse coding with average region partition is proposed. Firstly, local invariant feature transform （SIFT） feature extraction was applied to the original image. Then, Laplacian regularization was added to the sparse coding method to encode the local features so that the similar features have similar code words and the feature vectors were fused by average region partition and max pooling. Finally, multi-class SVM classifier was used to classify the images. Experimental results on several standard image datasets show that the algorithm has higher classification accuracy.