Current Issue Cover
  • 发布时间:
  • 摘要点击次数:
  • 全文下载次数:
  • DOI:
  • 2017 | Volume  | Number 6

()

摘 要
目的:针对稀疏编码方法的编码不稳定以及图像表示和分类相互独立的问题,提出非负局部Laplacian稀疏编码和上下文信息的图像分类算法。方法:利用非负局部的Laplacian稀疏编码方法对局部特征进行编码,并通过最大值融合得到原始的图像表示,有效改善编码的不稳定性。在所有图像表示中随机选择部分图像生成基于上下文信息的联合空间,并通过分类器将图像映射到这些空间中。将映射后的特征表示作为最终的图像表示,解决图像表示与分类相互独立的问题,使得图像特征之间的上下文信息更多地被保留。结果:在四个公共的图像数据集Corel-10、Scene-15、Caltech-101以及Caltech-256上进行了仿真实验,并和目前与稀疏编码相关的算法进行实验对比,分类准确率提高了3%~16%。结论:提出的非负局部Laplacian稀疏编码和上下文信息的图像分类算法,改善了编码的不稳定性并保留了特征之间的相互依赖性。实验结果表明,该算法与已有算法相比的分类效果更好。另外,该方法也适用于图像分割、标注以及检索等计算机视觉领域的应用。
关键词

()

Abstract
Objective:Sparse coding(SC) is unstable in the process of encoding, besides image representation and classification are relatively independent in SC method, a new method is proposed, called image classification with non-negative and local Laplacian sparse coding and context information (NLLSC-CI). Method:Firstly, non-negativity and locality constrained Laplacian sparse coding (NLLSC) is used to encode local features in order to obtain original image representation by max pooling (MP), which improves the instability of encoding effectively. Some image representation that is randomly selected from all images is generated joint spaces based on context information, and all images are mapped into these spaces by SVM classifier. The mapped feature representation is regarded as final image representation, which solves the problem that image representation and classification are relatively independent and preserves more context information between features from images. Result:In order to verify the efficiency of the proposed algorithm, four public image datasets including Corel-10, Scene-15, Caltech-101 and Calthch-256 are used to experiment. The results are compared with state-of-the-art sparse coding algorithms, which have suggested the classification accuracy increases 3%~16%. Conclusion:The proposed algorithm improves the instability of coding and preserves the mutual dependency between local features. Experimental results have shown that our algorithm has better performance than some previous algorithms have been developed. In addition, the novel method has proved capable of application to numerous computer vision such as image segmentation, image annotation and image retrieval.
Keywords
QQ在线


关注微信