Q91 : Content-baxsed Image Retrieval Using High Level Semantics
Thesis > Central Library of Shahrood University > Computer Engineering > MSc > 2016
Authors:
Abstarct: content baxsed image retrieval(CBIR) is one of the well-known fields of study in computer vision and image processing. A CBIR system searches and retrieves images from an image data baxse using their visual content. Due to the rapid increase in collections of images with high volume and lack of accountability text-baxsed image retrieval systems, content-baxsed image retrieval systems are essential. Current CBIR systems, mostly use only low-level features(such as color, shape and texture) of images. Low-level features can describe the content of easy images correctly, but can not describe complex images, including high-level concepts. A high-level concepts are concepts of human perceptual in image that are not understood using low-level features. This problem, is one of the main challenges of CBIR systems and called "semantic gap".
In this thesis, two methods are presented to reduce the "semantic gap". These two methods are following semantic information of the image using combination of low-level features. According to studies, the human visual system considers edge and color features in image evaluation. So these two features are covered semantic information and have a significant role in content-baxsed image retrieval systems. One of the methods in this field, often use edge features and color information in the L*a*b* color space. The feature vector include color differences baxsed on these features, is called "histogram color difference". In this thesis, first, by applying two criterias, the entropy and the correlation between feature vector components in different images, effective features are selected. Although these features represent the semantic content of the image partly, by adding new features to the feature set, rate of image retrieval can be increased. One of the proposed features, is the color histogram in the HSV color space. The HSV color space and L*a*b*, compared to other color spaces are closer to the human visual system. The L*a*b* color space is used in color difference histogram method. So, in the first proposed method, to use of semantic information in both color spaces, color histogram in the HSV color space is added to the feature vector. This feature vector in terms of features, speed and accuracy of image retrieval, compared to the newest methods, shows a significant improvement. The second proposed method follow semantic information using both edges, include the marginal edges of objects and internal edges of objects. The image texture including the internal edges of objects in the image. So, in the second proposed method, edge orientation histogram using gradient of image texture to add the feature set of first proposed method. Finally in criteria similar, to each group features of the proposed features, the weight is assigned. The assignment of these weights, is increased speed and accuracy in retrieval of proposed systems. The extracted semantic features with these appropriate weights, known as high levels semantic. Because leading to understand the high-level concepts of image. proposed methods improve rate of compared to recent other methods. In this thesis, the two standard databaxses Corel 5k and Corel 10k are used. This increases the accuracy and speed retrieval, demonstrate that proposed high levels of semantic, despite few features, without using of image segmentation and learning and clustering, describe two-dimensional image space well and semantic concepts are extracted.
Keywords:
#content-besed Image retrieval #color histogram #edge orientation histogram #entropy #correlation #high levels semantic #local binary patterns Link
Keeping place: Central Library of Shahrood University
Visitor:
Keeping place: Central Library of Shahrood University
Visitor: