多特征编码融合的图像分类研究

Multiple Feature-Coding Fusing for Image Classification

词袋(Bag-of-Words)模型是图像分类研究中使用最广泛同时也是最有效的框架模型之一。然而,字典的最优设计仍然是该模型的重要研究内容。直观来说,字典越大,图像分类的准确度就越高,但同时也需要更高的计算资源和存储代价。鉴于此,本文提出一种基于多特征融合的图像分类方法。首先提取并使用图像的视觉特征生成多个小字典,接着构建不同字典下的直方图交核以获得图像的特征编码,最后通过对上述编码进行在线学习加权融合,使得组合结果与大字典下获取的特征编码一样具有较强的判别性,从而提高图像分类准确度。在特征融合阶段,本文改进了OPA(Online Passive-Aggressive)算法,得到了权值更新的闭式解。实验结果表明本文方法运行效果良好且计算代价更低。

The Bag-of-Words model is one of the most popular and effective image classification frameworks in recent literature.However,the optimal formation of visual vocabulary remains an important research content.Intuitively,the larger size of vocabulary leads to higher accuracy,but which means that more computational resources and memories are required.In this paper,we propose a multiple feature-coding method for image classification.Firstly multiple small-sized vocabularies are generalized using extracted visual features of images.And then,histogram intersection kernels are constructed respectively under each vocabulary,which subsequently output as feature coding.Finally,an online weighted fusing method is adopted to obtain discriminative performance similar to large-sized vocabularies.Specifically,the weights are learned via an improved Online Passive-Aggressive algorithm in feature fusing stage,which leads to a closed-form solution.Experimental results on different image sets demonstrate that our proposed method performs favorably with lower cost against several typical methods.