童鞋论文网专业提供代写本科论文、代写硕士论文、代写职称论文、代写毕业论文、论文发表等服务,上万成功案例,欢迎咨询。 网站地图  |  加入收藏  |  设为首页
童鞋论文网
热门搜索
行政 教育 营销 法律 经济 护理 心理 会计 文学 工商 计算机 旅游
站内搜索:
童鞋论文网
 当前位置:童鞋论文网 > 硕士论文 > 文章正文
 
图像句子标注方法研究
论文作者:童鞋论文网  论文来源:www.txlunwenw.com  发布时间:2019/11/1 9:13:50  

摘要:图像句子标注是指利用计算机为图像生成一段连续文本,该文本能够准确描述图像中的关键视觉内容。图像句子标注属于机器视觉与自然语言处理的交叉学科研究,即通过机器视觉方法得到能描述图像的单词集合,再通过自然语言处理方法将单词集合生成一条准确描述图像的句子。

首先,图像句子标注利用机器视觉方法建立一个从图像底层特征到高级语义特征的映射模型,从而减小“语义鸿沟”。其次,利用自然语言处理方法将标注关键词转化为标注句子,为图像检索提供更详细、更具体的推荐,从而缩短了用户获取理想图像的耗时。图像句子标注将机器视觉与自然语言处理的研究较好结合起来,推动学科间的交叉、渗透,提出新的思路,推动理论研究的进步。

图像句子标注的研究极具挑战性,目前依旧存在特征学习不足和句子可读性不好等问题。故本文从以下两方面展开研究:

(1)特征学习是图像句子标注工作的前提,在机器视觉阶段,本文提出基于特征融合的图像句子标注。首先,详细说明了本文在机器视觉阶段的算法设计,本文以特征融合的方法提高机器视觉阶段的识别效果,用主成分分析法和特征归一化降低特征维数过高、计算量大的影响。最后设计与单特征、双特征的对比实验,并用查全率、查准率和算法耗时对实验结果评判,结果表明,本文融合算法在查全率和查准率都有提高。

(2)句子的可读性、准确性是图像句子标注工作的关键,在自然语言处理阶段,本文提出基于Sentence-Rank算法的图像句子标注。首先详细说明了本文在自然语言阶段的详细算法设计,说明本文标注模型的目的是找到一条准确、简洁并常用的句子来描述图像内容,利用句子筛选方法找到备选句子,然后用句子评分方法对句子进行排序,最后将评分最高作为图像句子标注。在实验阶段,本文用BLEU模型确定Sentence-Rank算法中N-gram模型的N值,然后用于与句子模板的对比实验,实验结果表明,本文算法平均困惑度更低,具有更好的可读性。

Image Sentence Annotation means usingcomputer as a center for generating a sequential paragraph for the image whichshould be able to indicate its vital visual contents directly and completely.To be more specific, as an interdisciplinary research of Computer Vision andNatural Language Processing. Image Sentence Annotation means compoundingnumerous words in a detailed sentence for the specified photo after acquiringeligible word-group from enormous database. Author here placed more emphasis onhow to promote the combination of two basic research fields.

Image Sentence Annotation sets up a mappingmodel from fundamental image features to advanced semantic features provided asolution of narrowing semantic gap. It uses national language progressing tolabel sentences rather than key words could promote the process of detailednessand accuracy of image retrieval. As a consequence, clients may have no needconsuming inconsequential time on searching for ideal images. Fresh thoughtsand innovative measures occurred during researches in binding Computer Vision andNatural Language Processing, which were considered as major roles in thepromotion of access to theoretical study.

The study of image sentence annotation isvery challenging. Nowadays, there are still problems such as lack of featurelearning and poor readability of sentences. Therefore, this article studiesfrom the following two aspects:

(1)Feature learning is the precondition ofimage sentence tagging. In the stage of machine vision, this paper proposes animage sentence annotation with feature fusion. In this paper, the featurefusion method is used to improve the recognition effect of the machine visionstage, and the principal component analysis method and feature normalizationare used to reduce the effect of the feature dimension being too high and thecalculation volume being large. The final design is compared with a singlefeature and a double feature, and the test results are judged by the recallrate, precision rate, and algorithm time-consuming. The results show that thefusion algorithm improves the recall rate and the precision rate.

(2)The readability and accuracy ofsentences are the key to image annotation work. In the natural languageprocessing stage, this paper proposes an image sentence annotation bases on theSentence-Rank algorithm. Firstly, the detailed algorithm design of this paperin the natural language stage is explained in detail. The purpose of theannotation model in this paper is to find an accurate, concise and commonlyused sentence to describe the image content, use the sentence screening methodto find the alternative sentence, and then use the sentence scoring method toThe sentences are sorted, and the highest score is finally marked as an imagesentence. At the experimental stage, this paper uses the BLEU model to determinethe N-value of the N-gram model in the Sentence-Rank algorithm. Then it is usedto compare with the sentence template. The experimental results show that theaverage perplexity of this algorithm is lower and it has better readability.

关键词:机器视觉;自然语言处理;特征融合;主成分分析法;Sentence-Rank

computer vision; national languageprogressing; feature combination; principal component analysis; Sentence-Rank

上一篇:三维石墨烯负载Pd催化剂的制备及加氢性能研究     下一篇:没有资料
 
论文客服 95702844
论文热线18796337551
本站网址 www.txlunwenw.com
郑重承诺 原创,包修改,包通过!
童鞋论文网真诚欢迎新老客户的光临与惠顾!
 
 热门文章
· 住宅建设用地使用权续期研究
· 用户锁定效应视角下的网易云音
· 众筹产品创新性对融资绩效的影
· 南京市城镇居民冷鲜肉认知度问
· 摩拜共享单车可持续盈利研究
· 催生地方政府隐性债务风险的原
· 社交媒体中不实信息转发主体心
· 基于网络平台的SPA模式研究——
· 基于游客感知的云台山景区微信
· 新媒体时代背景下乡村旅游企业
 
 推荐文章
· 广西地区农家乐经营模式研究
· 碳金融和绿色经济发展研究综述
· 信息经济学下的“考研占座热”
· 基于二手物品交易平台的可持续
· 论取消进口预包装食品标签备案
· 部编版小学语文文言文类文章编
· F模型下“嘉欣丝绸”财务风险研
· 基于4V理论的云南嘉华饼屋营销
· 论商业性小屏直播的主持互动策
· 新兴偶像经济发展现状及发展建
 
 相关文章
· 基于SPSS的主成分分析法在松花
· 非国家行为体的影响及其限度研
· 雷蒙德·威廉斯的生态社会主义
· 麦金太尔正义理论思想渊源研究
· 基于主成分分析法的我国海洋战
· 论做好硕士研究生的引路人
· 新时期研究生学术活动研究
· “石发”与文学创作之关系研究
· 我国男装奢侈品消费市场探讨
· 关于张家山汉简所显示的汉初法

论文原创,准时交稿,包修改,包通过,八年信誉,品质保证
童鞋论文网提供代写本科论文,代写专科论文,代写毕业论文,代写职称论文,代写硕士论文,论文发表及查重服务
备案号:苏ICP备11011187号-1 苏公网安备32030502000117号
版权归童鞋论文网所有 禁止一切商用盗用