基于空间信息和迁移学习的图像多标记算法研究

发布时间：2018-02-03 14:19

本文关键词： 多标记学习空间信息残缺图像关联性迁移学习　出处：《山东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：伴随着网络信息技术的飞速发展,互联网+模式的迅速兴起,人们对于网络信息的获取与需求呈指数般增长。除了对文字信息的需求外,对于图像内容信息的认知与理解也逐渐为人们所重视。图像自动标注技术的出现,在一定程上弥补了人工标注存在的耗时耗力、较为主观等不足,提升了图像理解技术的效率。但现今人们对于图像内容的理解已经不仅仅拘泥于单一的概念和标记了,而更倾向于多层次多角度的解读,图像的多标记学习应运而生,更好的适应了人们的需求。图像的多标记学习方法层出不穷且渐趋成熟,对于图像区域空间信息的运用也越来越充分,但是现实世界中除了完整的图像外还存在着大量残缺或者被遮挡的图像,其中也包含着大量有效的信息,针对这部分特殊的图像族群,运用空间信息,提出残缺图像的多标记学习方法,该方法可以减弱图像缺损部分对图像内容理解的影响,提高残缺图像的标注查全和查准率,更好体现整幅图像的蕴含信息。同时,图像的多标记学习中图像与图像之间,图像与标记之间,标记与标记之间的关联性还需更充分的利用,将机器学习中的相似性迁移思想融合进图像的多标记学习中,提出基于相似性迁移学习的图像多标记算法,探究图像与标记之间的关联性,能够有效提高图像的标注质量,减少噪声干扰。本文的主要工作与创新点概括如下:1.结合空间信息对于图像内容理解的重要性,针对残缺图像族群,提出一种基于空间信息的多标记算法。首先选取图像缺损部分的最小矩形局域,沿矩形边沿延伸将所有图像按此比例进行分割,然后以图像的分割子块为单位进行图像的相似性度量,利用图像分割区域的空间结构信息完成对图像的自动标注。这种方法能够充分的利用残缺图像的空间信息,减弱图像缺损部分对图像内容理解的影响,提高残缺图像的标注查全和查准率,更好的体现整幅图像的蕴含信息。2.为了进一步探究图像标记之间的关联性,融合迁移学习理论,提出一种基于相似性迁移学习的图像多标记算法。首先建立图像间的特征相似度量,然后引入相似性迁移学习算法,将图像的底层特征相似度量迁移到图像所对应标注词的相似度量,通过统计方法实现图像的自动标注。该方法能够有效提高图像的标注质量,减少噪声干扰,为图像多标记学习提供额外的有用信息,在一定程度上弥补了样本数据的不足。通过运用图像空间的区域结构信息,融合迁移学习理论将图像相似性迁移到图像的标记学习中,论文中的多标记学习算法对于残缺图像族群能够有效提高其标记性能,具有良好的鲁棒性;对于完整图像族群,可以有效弱化干扰,增强其标记学习效果。
[Abstract]:With the rapid development of network information technology and the rapid rise of Internet mode, people's access to and demand for network information has increased exponentially, in addition to the demand for text information. Recognition and understanding of image content information has gradually been paid attention to. The appearance of automatic image tagging technology, in a certain process, make up for the shortcomings of manual annotation, such as time-consuming, more subjective and so on. It improves the efficiency of image understanding technology. But nowadays, people's understanding of image content is not only limited to a single concept and label, but also more inclined to multi-level and multi-angle interpretation. Image multi-label learning emerged as the times require to better meet the needs of the people. Image multi-label learning methods emerge one after another and gradually mature, the use of spatial information in the image region is becoming more and more fully. But in the real world in addition to the complete image there are also a large number of incomplete or occluded images which also contain a large number of effective information for this part of the special image groups the use of spatial information. This paper proposes a multi-label learning method for incomplete images, which can reduce the effect of image defects on image content understanding and improve the tagging and checking accuracy of incomplete images. At the same time, the relationship between image and image, between image and label, between mark and label should be used more fully in image multi-label learning. The idea of similarity transfer in machine learning is integrated into image multi-label learning, and an image multi-label algorithm based on similarity transfer learning is proposed to explore the correlation between image and label. The main work and innovation of this paper are summarized as follows: 1. Combining the importance of spatial information for image content understanding, aiming at incomplete image groups. A multi-label algorithm based on spatial information is proposed. Firstly, the minimum rectangular region of the defective part of the image is selected and all images are segmented in this proportion along the edge of the rectangle. Then we measure the similarity of the image in the unit of segmentation sub-block. This method can make full use of the spatial information of the incomplete image and reduce the influence of the image defect on the understanding of the image content by using the spatial structure information of the image segmentation region. In order to further explore the relevance of image markers, fusion transfer learning theory is used to improve the tagging and precision rate of incomplete images, and better reflect the information contained in the whole image. 2. An image multi-label algorithm based on similarity transfer learning is proposed. Firstly, the feature similarity between images is established, and then the similarity transfer learning algorithm is introduced. The image's bottom feature similarity is transferred to the image's corresponding tagged word's similarity, and the image's automatic annotation is realized by statistical method. This method can effectively improve the image's tagging quality and reduce the noise interference. To provide additional useful information for image multi-label learning, to a certain extent, to make up for the lack of sample data, by using the image space of regional structure information. Fusion transfer learning theory transfers image similarity to image tagging learning. The multi-label learning algorithm in this paper can effectively improve the marking performance of incomplete image populations and has good robustness. For the complete image population, the interference can be weakened effectively and the effect of marker learning can be enhanced.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】