基于核拉普拉斯稀疏编码模型的图像分类

Sparse Coding Model Based on Kernel Laplacian for Image Classification

在稀疏词袋模型中,由于码书的过完备性,相似特征间稀疏编码存在多种组合方式,从而导致相似的特征可能得到完全不同的编码.文中提出基于核拉普拉斯稀疏编码的图像分类方法.该方法首先通过在高维核空间中构造核拉普拉斯矩阵以描述特征间的几何依赖关系,使相似特征的稀疏编码的相似性尽可能得到保持.再采用交替固定码书与稀疏矩阵的方法优化目标函数进行码书学习,并采用符号猜测法对特征进行稀疏编码.最后用多类SVM分类器分类.实验证明文中方法可较大幅度降低量化误差,提高分类准确率,并在多个数据集上取得良好的测试效果.

In bag-of-words with sparse coding model, similar features can be encoded as various sparse coding combinations due to the over-completeness of the codebook, which results in totally different visual words. In this paper, a sparse coding method based on kernel Laplacian for image classification is proposed. Firstly, a Laplacian matrix is constructed to capture geometric dependencies between the features in high-dimensional kernel space, and thus the similarity of sparse coding between the similar features can be maximally preserved. Secondly, the objective function is optimized for codebook learning by fixing codebook and sparse matrix alternately, and feature-sign search algorithm is used for sparse coding of the features. Finally, the one-to-all linear SVM classifier is applied to classify images. The experimental results on several datasets show the proposed algorithm decreases the quantization error dramatically and improves the classification performance.