Research On Some Key Techniques Of Content-based Image Retrieval

With Digital Technology the rapid development of digital cameras, video cameras and other digital equipment, flying into the homes of ordinary people. A variety of multimedia data such as images, video and other rapidly expanding gradually become the information in the mainstream, and people's lives have an important impact. As the image data rendering geometric progression of growth, the need for images has become in recent years, image retrieval on the main hot spot. Meanwhile, Image retrieval techniques, but also caused the academic and industry attention.Study of visual perception of the retrieval and modeling, and analysis of their learning mechanism, and explore in line with the human perception of visual media interaction, has become the image retrieval as an important area. From nineties began to people from the visual aspects of the image features in-depth study. Colors relative to the geometric shape, the right translation, rotation, scaling transformation with invariance demonstrated considerable robustness. We passed on several common color modeling principles for discussion and analysis, and drawing on cognitive psychology research, presented with the human visual perception consistent with HSV quantitative model, on this basis is given based on HSV Model Image Retrieval algorithm, effectively reduces the complexity of the algorithm to improve the retrieval efficiency.In the image retrieval process, the introduction of machine learning methods can make use of the existing machine learning field theory of the learning process research and analysis. The data-based machine learning of modern intelligence technology in the important aspects. Research from the observed sample off to find the laws, the use of these laws for the future of data or not observed data to predict. Including pattern recognition, neural network, etc., the existing machine learning methods based on shared premises is one of statistics, the traditional statistics is the study of the number of samples tends to infinity when the incremental theory of theory, the existing learning methods are mostly based on the this assumption. Based on image feature extraction method can greatly promote the content-based image retrieval research, it is based on the image's own content. However, due to the image low-level features and high-level semantics between the semantic gap, can not use the image low-level feature accurately convey the image of the high-level semantics. In order to effectively address the outstanding issues, on the one hand the need in image processing research in the field of effective feature representation method, on the other hand is an attempt to capture and build low-level features and high-level semantic link between, is how to narrow the semantic gap problem.Support vector machines in recent years in the statistical learning theory developed on the basis of machine learning methods, this paper, the SVM method of study will SVM for image retrieval, interactive learning process, the user feedback and relevant information as the two classification of training samples, by learning to be two classifiers, and then the entire image retrieval is proposed based on support vector machine SVM interactive image retrieval algorithm. Experimental results show that, SVM method under ideal conditions to achieve a better than other machine learning methods of search results. Kernel function selection problem is usually in the SVM method is a more difficult problem, the optimal kernel function selection has no theoretical methodological guidance, the same, covering the problem is also a NP problem, not solving the optimal solution, it is still not certain the theory to guide. The use of nuclear function, so that the sample is mapped to a high-dimensional space, and then construct the optimal separating surface. Therefore, kernel function selection directly affects the classifier's generalization ability.Kernel function selection, including nuclear functions and parameters to determine the two aspects. Through the experimental comparison concludes that Gaussian kernel function in the image retrieval learning the best, so we choose Gaussian kernel function as kernel function, and then Gaussian kernel function parameters by empirical means to select and defined. Experiments show that illustrates the SVM in the first training sample cases, still be able to maintain a good learning ability and generalization ability.Relevance feedback technique is an interactive search technology, allowing users to retrieve the results marked by learning user feedback and relevant information to improve search results. In this paper, image retrieval in this interactive learning process, and draw on the relevant feedback technique is proposed based on radial basis function to the neural network learning algorithm for interactive image retrieval methods, the use of multiple neurons covered by the geometric area of the combination of sample data to describe the effective representation of the relevant images in the feature space distribution. The experimental results show that this method can effectively improve the search results to make it more in line with the needs of users. As the neural network and the human brain has a similar mechanism, therefore, for large-capacity parallel image retrieval with with a clear advantage, and achieved good results.In order to overcome the semantic gap of the impact of this paper, semantic-based technology research in this area. Images contain rich semantics, in the semantic image retrieval research occupies an important position. Semantic description, feature extraction, recognition is semantic retrieval of the most central issue. Presented based on image characteristics of semantic retrieval, which made up based on visual features of the shortage reflects the people-centered concept. Based on the analysis semantic retrieval context, analysis monitoring semantic annotation and unsupervised semantic annotation. Are given based on semantic space, the image method is based on the user's implicit semantic features. And to achieve a semantic image retrieval method, which in fact based on the semantic space for retrieval of a special case, will the user's implicit semantic feature explicitly for explicit semantics. The image contains rich semantics, semantics as an important high-level semantic content, semantic image retrieval research occupies an important position. Semantic description, feature extraction, recognition is the emotional semantic retrieval of the most central issues. This in-depth study of these issues was proposed based on a new image content description framework, and implemented an image recognition prototype system. The results show that the prototype has achieved good results. Image Retrieval as a challenging research and a project of application, have broad application prospects. At the same time as time goes on, image retrieval from the laboratory to the real practical use would be just around the corner.