Abstract: In the traditional image semantic sentence annotation, sentence templates are used to describe the content of image. However, it is hard to meet the logic of language using the traditional method. Aiming at this problem, this paper proposes to describe the image content by selecting an optimal sentence from the corpus , and design the Sentence-Rank algorithm with the [N]-gram algorithm to generate the annotated sentence. Firstly, the HSV-LBP-HOG fusion feature with the best performance is used for image classification, the image markings are obtained. Then, it uses the string matching algorithm to list all the sentences with the marked keywords from the corpus and sorts the obtained sentences by Sentence-Rank algorithm, and selects the highest rated sentence to describe the image. The experimental results show that the annotation sentence obtained by this method has lower perplexity, and solves the linguistic logic problem of sentences better.

Key words: machine learning, natural language processing, feature fusion, Sentence-Rank, [N]-gram

摘要： 传统的图像语义句子标注是利用句子模板完成对图像内容描述，但其标注句子很难做到符合语言逻辑。针对这一问题，提出基于统计思想从语料库中选出一条最优的句子来描述图像内容，设计以[N]-gram算法为主要思想的Sentence-Rank算法生成标注句子。首先执行机器视觉特征学习，选择标注性能最好的HSV-LBP-HOG融合特征完成图像分类，获得图像标注关键词。然后，利用字符串匹配算法从语料库中列出包含所有标注关键词的句子，并将得到的句子通过Sentence-Rank算法进行价值排序，选取评分最高的句子描述图像。实验结果表明，该方法得到的标注句子具有较低的困惑度，较好地解决了句子的语言逻辑问题。

关键词: 机器学习, 自然语言处理, 特征融合, Sentence-Rank, [N]-gram