Interactive Image Retrieval Based On Content Perception

With the wide use of digital image in the Internet and Multimedia technology, one of the most important issues in image processing, machine vision and information searching field has revealed itself on achieving more fast and effective ways of performing massive image data search. Through extracting low-level visual features of images, indexing is constructed by content-based image retrieval (CBIR), which is now the main way in image retrieval technology. Image content is revealed to some extent by the use of low-level visual features, however, taking account the subjective nature of the visual sensing of a certain human, a research based on content perception is needed to further satisfy the user’s retrieval intention. The main focus of this research is the reduction of the semantic gap between image low-level features and high-level semantics. To achieve the goal, two interactive image retrieval algorithms are proposed in this paper, which combine the images’ visual features with users’ feedback information.First, a MLRI-based image retrieval framework is proposed based.on image’s statistical information and users’ feedback information. Innovation points:(1) A statistical index called visual region importance (RI) is constructed, and it can be updated as semantic RI according to users’ feedback information, this will reduce user’s burden in selecting interesting regions. (2) In order to make full use of history retrieval information, a memory learning RI (MLRI) technique for region-based image retrieval (RBIR) system is developed, which can tap the semantic information of images fully and update adaptively by studying the RI in a long-term. (3) Specially, through performing region clustering based on optimal query, the MLRI-based image retrieval emphasizes the latest positive regions by assigning bigger weights to them. Experiments show that the proposed MLRI-based RBIR algorithm is effective and robustness while making full use of history retrieval information.Second, in CBIR, a PSP-based interactive image retrieval technology is proposed, which addresses the issue of query vector optimization constructed by query point movement (QPM) algorithm. Innovation points:(1) As particle swarm optimization (PSO) is only suitable for the functions predefined, this paper presents an improved PSO approach named particle swarm programming (PSP). (2) In order to provide a convenient way of representing arbitrary computational procedures, PSP changes the particles’internal structure dynamically by a hierarchical tree structural representation, which can increase the diversified of variables in problem space. (3) Specially, a nonlinear updated query vector is constructed by it opening a nonlinear version of updated query vector, and it can satisfy different users’ retrieval intention by assigning weights for feedback images adaptively. Taking the overall proportion of the different positives’and negatives’ into consideration, the weights of annotation images are changed dynamically. Experiments for different categories on Corel image database verify the feasibility and effectiveness of the proposed framework. In the comparison with other states-of-the-art methods based on swarm optimization algorithm, the feasibility and effectiveness of the new method has been explained.