Similarity-based image retrieval method using wavelet transformation and its application to image database of historical objects in Japan

本课题的目的是为基于内容的图像数据库检索提供一种基于小波变换的统一方案。该方案的两个主要组成部分如下：1)利用正交离散小波变换(DWT)对图像进行分层分解：将一幅图像分解为三幅方向选择性的细节图像和一幅近似图像。分解过程可以递归地应用于近似图像以产生下一级别的分辨率。这样，我们就可以得到对应于不同尺度的不同分辨率的子图像的金字塔结构。需要注意的是，金字塔结构是图像处理的基础，如边缘提取、纹理提取、纹理分割、形状提取等。2)特征提取：我们从离散小波变换得到的子图像中提取图像的特征，例如，我们使用高阶局部自相关作为特征来检索图像、形状、用户构造的草图和图纸。对于纹理检索，我们提出了由细节图像的小波系数计算的新特征。基于这一方案，我们开发了日本各地历史文物图像数据库的原型系统。该数据库包括日本国家历史博物馆收藏的经典图片、图案和工具，它们是日本遗产和/或遗产物品。该系统允许基于示例图像、用户构建的草图和图纸、纹理以及这些的交互组合进行相似性检索。然后，系统支持浏览检索，当用户对自己的检索需求不明确或不熟悉数据库中的可用信息时，需要通过浏览进行检索。对数百幅图像进行的实验表明，该算法具有较好的性能。

The aim of this project is to provide a unifying scheme based on wavelet transform for content-based retrieval of image database. Two main components of the scheme are as follows :1)Hierarchical decomposition of images using orthogonal discrete wavelet transform (DWT) : DWT decomposes an image into three orientation selective detail images and an approximate image. The decomposition process can recursively be applied to the approximate image to produce the next level of the resolution. Thus we can obtain a pyramid structure of subimages with various resolutions corresponding to the different scales. It should be noted that the pyramid structure is the basis for image processing such as edge extraction, texture extraction, texture segmentation, shape extraction, and so on.2)Feature extraction : We extract features of an image from the subimages obtained by DWT.For example, we use the higher-order local autocorrelation as features for retrieval by example image, shape, user-constructed sketches and drawings. For retrieval by texture, we proposed new features computed from the wavelet coefficients of detailed subimages. These features are stored in the database.Based on this scheme, we have developed a prototype system of an image database of the historical objects of various places in Japan. This database includes classical pictures, patterns, and instruments, which are Japanese heritages and/or legacy objects collected in the National Museum of Japanese History.The system allows similarity retrieval based on example images, user-constructed sketches and drawings, textures, and interactive combinations of these. Then the system supports retrieval by browsing which is needed when users are vague about their retrieval needs or unfamiliar with the information available in the database. Experiment with hundreds of images showed good performance.