一种基于稀疏编码空间金字塔匹配的图像分类算法

Sparse coding based spatial pyramid matching algorithm for image classification

图像分类技术是近年来计算机视觉领域中的研究热点,在移动互联网领域中取得了成功应用。提出了一种基于稀疏编码空间金字塔匹配的图像分类算法。该方法首先对图像的SIFT特征进行稀疏编码,替代了传统的矢量量化方法,可以有效降低量化误差,构建更为准确的图像表征方式,然后结合空间金字塔匹配算法采用线性分类器对图像进行分类识别。在标准测试图像数据库上的实验结果表明,相比BOF和SPM方法,该算法可以将图像分类准确率提高4%~12%。

Image classification is a hot issue in computer vision field and has been extremely popular in mobile internet applications. A sparse coding based spatial pyramid matching（SCSPM） algorithm was proposed. Firstly, the extracted scale-invariant feature transform（SIFT） descriptors are encoded by sparse coding method instead of the traditional vector quantization method. The sparse coding step can reduce the quantization errors effectively and generate more discriminative image representation. Furthermore, image classification can be obtained by linear spatial pyramid matching method. The experimental results on Caltech 101, Caltech 256 and 15 Scenes data sets show that our method can significantly improve the image classification accuracy by 4% -12% compared with the bag-of-featwes（BOF） and SPM algorithms.