基于共生局部特征的图像检索研究

In recent years, many content-based image retrieval methods extract local features, and then organize the spatial relationship to perform similarity search. However, because there is no inherent relationship between the extracted local features, there exists uncertainty and randomness when organizing the spatial relationship. Furthermore, the spatial relationship is not invariant to translation, rotation and scale at the same time. To enhance the robustness of spatial relationship, the complexity of the retrieval methods should be increased, but the retrieval efficiency is decreased. Thus the retrieval efficiency should be improved by sacrificing some robustness of spatial relationship. To solve this problem, we introduce the problem of organizing spatial relationship into local feature extraction, and propose the concept of the co-occurring local features. The related retrieval method is detailed as follows: First, the spatial relationship is explored in the process of extracting co-occurring local features, with the purpose of being invariant to translation, rotation and scale simultaneously, and decreasing the complexity of retrieval method. Then, to improve the spatial similarity between images, the co-occurring local features are selected by evaluating the spatial information, and the spatial relationship is introduced into the co-occurring local features quantization. Finally, based on the quantized co-occurring local features, the multi-dimensional inverted index which contains the spatial relationship is generated, and the related spatial similarity measurement is proposed to improve the retrieval performance.

近年来，基于内容的图像检索算法大多先提取局部特征，再组织局部特征的空间关系进行相似性搜索。然而，由于局部特征之间不存在内在联系，导致组织空间关系时存在随机性和不确定性，进而难以同时保证平移、旋转和尺度不变性。为了增强空间关系的稳健性，需要提高算法复杂度，但将降低检索效率，从而需要牺牲空间关系的部分稳健性来换取检索实时性。为此，本项目将如何组织局部特征的空间关系这一问题，上升到特征提取层面进行研究，提出共生局部特征的概念，并进行如下工作：首先，在共生局部特征提取过程中探究局部特征的空间关系，达到既不增加算法复杂度，又可同时保证平移、旋转和尺度不变性的目的。接着，通过判别空间信息量筛选共生局部特征，并将空间关系引入共生局部特征的量化过程，期望增强图像之间的空间相似度。最后，基于量化后的共生局部特征创建包含空间关系的多维倒排索引，并提出相应的空间相似性度量方法，以便提高检索性能。

近年来，随着图像的数量爆炸式地增长，基于内容的图像检索方法成为人工智能领域研究的热点之一。本项目针对图像检索中特征提取以及相似性搜索问题，以共生局部特征及其空间关系作为主要研究对象，以深度学习算法作为辅助研究对象，在图像检索研究的基础上延伸到视频检索研究，主要完成了以下工作：(1)提出了基于共生局部特征的图像检索方法，并将其应用到杂志封面图像检索中，在保证检索效率的同时提高了检索准确率；(2)提出了基于自适应矩形窗口的图像检索方法，通过组织局部特征的空间关系来提高检索准确率；(3)提出了基于显著性区域的视频关键帧提取方法，将视频检索转化为图像检索，在安防监控视频中验证了基于共生局部特征的图像检索算法的可行性；(4)在深度学习算法的基础上，对图像检索、图像配准和图像识别等进行了初步研究和探索。本项目的研究成果预期能够在景点旅游、网上购物、医学诊断、安防监控等领域得到应用。