Research On Example And Sketch Based Image Retrieval

As the development of multimedia information technology and Intemet technology, image database is getting larger and larger. How to find a special image of interest from the mass database efficiently and effeetively has become a major issue to be solved. In this case, the research on content-based image retrieval was developed, and became a hot spot in the fields of multimedia information processing. The most existing systems of CBIR fulfill the tasks of retrieving the similar images through computing the degree of similarity of different images. Though those methods have achieved some success, the retrieval results through most existing methods are still not satisfactory. The main difficulty is that, the qualities of retrieval results by computers are much dependent on the features of images, and they have great differences with human beings who predict the degree of similarity of images through the image semantics. So, how to automatically acquire image semantics which meets the needs of users is the key to solving the problem of image/video retrieval.To solve this problem, this dissertation focuses on the research of image semantics extraction and user semantics acquisition. The main content and contribution are summarized as below:Firstly, this dissertation introduces the study background and significance of image/video retrieval, and provides a survey of the most important issues in the current literature in image and video indexing and retrieval.Subsequently, a novel semantic object extraction method——husking algorithm is presented in this dissertation. Firstly a Luv color histogram of image based method is used to estimate the color bandwidth, and then a mean shift algorithm with adaptive color bandwidth is employed to segment the inputted image. Finally, a husking algorithm is proposed to extract semantic objects by stripping away the background regions from the segmented image step by step. The experimental results show that the husking algorithm is an effective unsupervised one which needn't the participation of user and can extract obvious semantic objects from an image.After that, this dissertation introduces a novel method for freehand sketch retrieval based on normalized shape. Firstly a Fourier-based approach for dimensionality reduction and smoothing is employed to process object contour data. And after normalization, the sampled contour data is used to extract contour and region features of object. And then a new affine adaptive skeletonization algorithm is presented to extract object skeleton. Then we present a new skeleton tree descriptor and matching algorithm. Ar last, a new sketch retrieval method combining contour region and skeleton features is presented. The retrieval method proposed uses not only the contour from outer and the region from inner of an object, but also the skeleton that preserves the original object's topology and shape. The experimental results show that the method is robust to the 2D object affine transformation (translation, scaling, rotation) and noise corruptions.Finally, this dissertation presents a novel query-by-sketch and relevance feedback based semantic object image retrieval system, aimed at reducing the semantic gap between low-level features and high-level semantics. Firstly the system retrieves semantic object images and their regions sets by calculating the distance between the shape features of the query sketch and the shape features in the object feature database. And then the system extracts user's semantic features from the query sketch and the object images selected by the user. Finally, the system accomplishes the retrieval task on the image feature database by using region matching method. The system can also uses an object image selected by user as a query example to do image retrieval. The experimental results show that the proposed method is effective in retrieving images and can mark semantic object in the image.