Similar Image and Video Retrieval Systems Based on Spatial and Color Attributes

• Main Authors: Yung-Kuan Chan, 詹永寬
• Other Authors: Chin-Chen Chang
• Format: Others
• Language: en_US
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/68710183283439231139

博士 === 國立中正大學 === 資訊工程研究所 === 89 === The spatial attribute is the most natural way to describe an image for human beings. However, generating the symbolic representation by automatically extracting objects from an original image is very difficult. The color attribute is a very important cue in extracting information from images; based on the color attribute, the database system can be implemented automatically without human intervention. The goal of this dissertation is to investigate the spatial and color information of images and to evolve an effective and efficient similar image retrieval system from both attributes above. Additionally, this dissertation still extends the techniques of the image retrieval to those of video retrieval. To accord with the requirements of different applications, this dissertation presents a spatial similar image retrieval method based on the concept of tolerable difference of direction. In this method, users can determine the tolerable difference of direction to satisfy their requirements by themselves. This dissertation also extends it to recognize rotational invariance images, in which it is needless to specify where the rotation center is. The relation of spatial and temporal plays an important role in indexing video information. This dissertation defines a 9DLT string to outline the spatial relationships among the objects of a symbolic image, and proposes a dynamic programming method to indicate the matching sequence of query frames and database frames. It also defines a common component binary tree (CCBT) to store a set of 9DLT strings. This structure cuts down not only a great portion of storage space, but also the computing time of image matching distance between query video frames and database video frames. Traditional methods obtain the minimal video frame sequence matching (VFSM) distance between two video frame sequences Q and V by calculating the image matching distances of all the possible frame pairs (qi, vj). Here, qi is the i-th frame of Q, and vj is the j-th frame of V. Obviously, it is a time-consuming strategy. This dissertation provides a piecewise method to overcome this drawback, which needs to compute the image matching distances of the frame pairs (qi, vj) only if qi is similar to vj. To speed up video retrieval, this dissertation is engaged in developing a fast filter to prune off a large number of unqualified videos. This filter uses an image signature to indicate the spatial information of the objects in a video frame. It also introduces a cutting method to split a video into segments, where the image signatures in a segment are similar. The proposed filter, requiring only little memory space, can quickly cut away a large number of unqualified videos. Color histogram is the most commonly used color feature in image retrieval techniques. However, this feature cannot effectively characterize an image since it captures only the global properties. To make the retrieval more accurate, this dissertation introduces the run-length feature. The feature integrates the color distribution and shape information of the objects in an image. It can effectively discriminate the directions, sizes, and geometrical shapes of objects. Yet, extracting the run-length feature is time-consuming. For this reason, this dissertation also provides a revision of the run-length, called semi-run-length. This dissertation also introduces three simple and effective image features ─ the color moment (CM), color variance of adjacent pixels (CVAP), and CM-CVAP. The CM feature delineates the color-spatial information of images, and the CVAP feature describes the color variance of pixels in an image. However, these two features can only characterize the content of images in different ways. This dissertation hence provides another feature CM-CVAP, which combines both features above, to raise the quality of similarity measure. This dissertation proposes another three image features as well ─ the regional color moment feature (RCM), the wavelet-based feature, and the RCM-wavelet feature, which are useful for retrieving color images from an image database. The RCM feature is a revised version of the CM feature. The wavelet-based feature can describe not only the local and global color variances of pixels, but also some geographic shapes of objects in an image. The RCM-wavelet feature integrates the RCM and wavelet-based features to raise the quality of similarity measure. To get the image retrieving more efficient, this dissertation still presents a fast simple filter in terms of the color and spatial distribution of images. Requiring little storage space, the proposed filter can prune away many unqualified images quickly.