In this paper, we propose a novel model to improve image classi?cation based on the bag of visual words (BOW) model[1][5]. Our goal is to achieve higher image classi?cation accuracy by ?nding out the more signi?cant words who is of signi?cant topics. Recently, Fei-Fei et al. [9] and Sivic et al. [10] have applied bag of words model to the visual domain and achieve better results. But this model has some serious drawbacks. One of the shortcomings of this representation is the presents of the noisy visual words due to the coarseness of the vocabulary building process. Pierre Tirilly[15] use the PLSA[2] model to ?nd semantic words who have signi?cant semantic topics. But the method has some serious drawbacks. Here, we use the new technique to eliminate useless words which is based on the use of Probabilistic Latent Topic Analysis (pLTA), which is extended from Latent Semantic Analysis(LSA)[16] and Expectation Maximum(EM)[17] algorithm. Here we call this method pLTA model. For comparison, we detail the normal BOW model results, the PLSA results and the PLTA results in our experiments. Experiments show that our techniques can signi?cantly ?nd the signi?cative words among all the computed visual words, and achieve higher accuracy of image classi?cation.
展开▼
机译：在本文中，我们提出了一种新颖的模型来改善图像分类？基于视觉词袋（弓）模型[1] [5]。我们的目标是实现更高的图像类别？阳离子准确性？Ndude of signi？不能言语？无法主题。最近，Fei-Fei等人。 [9]和sivic等人。 [10]将单词模型应用于视觉域并达到更好的结果。但这种模式有一些严重的缺点。这种代表性的缺点之一是由于词汇建设过程的粗糙度而嘈杂的视觉词。 Pierre Tirilly [15]使用PLSA [2]型号到？ND语义单词谁有语义主题。但该方法具有一些严重的缺点。在这里，我们使用新的技术来消除基于概率潜在题目分析（PLTA）的无用词语，这与潜在语义分析（LSA）[16]和期望最大（EM）[17]算法延伸。在这里，我们称之为PLTA模型。为了比较，我们详细介绍了正常的弓形模型结果，PLSA结果和PLTA导致我们的实验。实验表明，我们的技术可以在展开？不知道的那种角度，在所有计算的视觉词中，实现了更高的图像类别的准确性吗？阳离子。
展开▼
