|
The computer intelligence has always been the research on artificial intelligence where the scholars and the experts on the Science of Computer are engaged in it and want to apply it to daily lives and practice. One of its researches is releated to Content-based image retrieval. Obviously, content-based image rertrieval has been explored and exploited since nineteen seventy’s, has transformed from the TBIR, CBIR to SBIR as a current research for more than decades and has achieved optimal evaluation and fruits on lab. For example, it advanced the MPEG-Seven standard, and designed the QBIC system. However, there has been a suspended but typical problem of how to analyze, to recognize, and to understand the image content through the computer. Especially, the "semantic gap" interrupts to bridge between the simple visual feature(like the image’s color, its shape, its texture, its spatial position, and so forth) and abundant semantics delivered by an image.This "semantic gap" make the process of content-based image retrieval towards the computer’s intelligence lag.For traditional "semantic gap" between the raw image features and the advanced semantics, subfunction unit done to retrieve the image is regared as a basic object. This paper defines the "object semantics", with object-oriented viewpoint and proper algorithms, explores and designs the process of object semantics-based image retrieval. In addition, this paper focuses on the crucial techniques of the image retrieval and has finished the main works as follow:1. This paper summarizes the development of color-based image retrieval, studys and adjusts efficient algorithms to extract color features. Meanwhile, this paper, on the basis of the supiority of color as the image feature, evaluates methods of description of the image color, like the simplest color histogram, the color moment, the color entropy, and so forth, studys and adjusts implements of the color quantification, such as adjusted image extraction with octree structure which is used to store the image color. After summarization of segmentation of the image, the image retrieval is relied on the colore feature to work out.2. For traditional "semantic gap" and its expression as "weak treatment of annotation", this paper fuses nerual network-like properties of artificial intelligence (such as sigmoid network) into automatic annotation of image semantics, traces back different segments of the same nerual network function which includes unique human abilities of the analysis and learning, and ensues to advance different treatments to be discussed as application. According to two words with relevent image content, or three and so forth to be classified, a multiple key words-based treatment of annotation is advanced, like D*={(I1,ω1,...,ωN),...,i(ID,ωD-N-1,...,ωD),],(?)N∈finited series3. Based on the new treatment of annotation and the same key words set with relevent contents, knowledge about Bayesian axiom and prior-statistics is learned once more, with increasing scale algorithm, and afterwards calculated to obtain the projections from the image feature X to key wordωi, or key words sets {ωi,ωj} within sets of description of the semantics {Li}, which impletes implicit relevent documents-based explicitly automatic anantation of the image key words.4. In course of the automatic annotation, This paper considers the complex of communicated languages between human beings (like the single word with multiple semantics, similar words), anatomizes traditional, linked closely, and undescribed structure between the description of the complicated senmantics and the extraction of image semantics, and designs oriented-objecet functions of image retrieval to implete the preprocess of semantic classification by using SVM.5. Although this paper has advanced annotation of the image with key words aiming at the image retrieval, which indeed can sovle the problem of indefinite words, this method with short term learning instead can not conquer subjective annotation of the image (for example, the person who pressed in the key word to descibe the image is not the same one who queries the same image by own key words). Thereforce, this paper has to trace and to use feed back results from the users after implement of the image annotation, and continues to bring forward SVM-based semantic solution of the image. This solution can grasp the real desire of the users and gradually comes close to the advanced definition linked with raw feature from the image. In algorithm to implement in detail, kernel widthσin kernel function and misjudgement parameter yi is precisely used as interface and interaction between users and the image retrieval. Just as interaction between users and the computer for many times, distinguishment between desire image of users and goal image through the image retrieval of the computer can be the minimum, and finally found as the correct image from the image retrieval.
|