The Application Of Visual Attributes To Content-Based Image Retrieval

With the progress of science and technology, the development of electronic equipment makes camera means more and more diverse. Camera equipment is at the fingertips. Network uploading and sharing technology leads the number of images on the internet growing rapidly. Effective image retrieval technology has become an urgent requirement. The traditional text-based image retrieval needs additional human-annotated labels. With the growth of the number of images, text-based image retrieval requires higher and higher human costs. In this trend, labeling images on the internet artificially is almost impossible in the near future. Content-based image retrieval is able to use the image content itself directly. Therefore, it is becoming a hot research field of image retrieval in recent years.Most of the content-based image retrieval methods use low-level features of the image to represent the image content and compare the degree of low-level features’ similarity to characterize the similarity degree of the image content, which brings the "semantic gap". Accurately speaking, the "semantic gap" refers to the difference between the computers to recognize the low-level features and the human beings to use the high-level semantics. To reduce the semantic gap, the image attribute is introduced into the field of image retrieval. Visual attribute is a semantic bridge, bridging the gap between high-level semantics and low-level features. This paper is expanded around the application of visual attributes to content-based image retrieval and does the following research:First, this paper introduces the applications of image visual attributes in the various fields of image processing. By analyzing those applications, it can be concluded that image attributes are being widely used in the field of image retrieval. Then, the four important attribute-based image retrieval algorithms, including independent attributes retrieval algorithm, dependent attributes retrieval algorithm, relative attributes retrieval algorithm, and weak attributes retrieval algorithm, are studied carefully in detail. This paper points out that the four algorithms tie closely, and analyzes the strengths and weaknesses of each of them theoretically and experimentally.Second, for the relative attributes retrieval algorithm using only the in-attribute class relationships, ignoring the between-attribute class relationships, this paper presents a structural relative attributes retrieval algorithm, drawing on dependent attributes retrieval algorithm and the structural support vector machine. By the comparison experiment of the five algorithms on three different types of databases (OSR, PubFig, and Shoes), the effectiveness of the proposed structural relative attributes algorithm is verified. Experiment results show that the proposed structural relative attributes retrieval algorithm is better than the best of the other four algorithms (relative attributes retrieval algorithm). It improves the AUC about3.86%in average.Third, all the current attribute-based image retrieval systems compel users to select attributes from the attribute set, which is extremely inconvenient when the attribute set is large. This paper presents an extended attributes retrieval algorithm, inspired by the weak attributes retrieval algorithm, via using a semantic relationship library named WordNet. The quantitative and qualitative comparison experiments are conducted on a-Pascal and a-Yahoo databases. Experiment results show that the proposed extended attributes retrieval algorithm gets a higher AUC on a-Pascal and a-Yahoo databases than weak attributes retrieval algorithm, which is the best algorithm on those two databases. The effectiveness of the system using extended attributes retrieval algorithm is better than the non-extended attributes system, increasing by30.4%.The subject comes from the major projects of the National Natural Science Foundation with Project ID:70890083.