Research On Some Techniques Of Content-based Image Retrieval

With the advent of the Information Society, the development and wide using of large memory storage device such as scanners, digital cameras etc, and the improvement of multimedia technology and the rapid growth of internet popularity have made image data grow explosively. How to find exactly the image which the users want in the vast image database has become multimedia research focus in the past decade.Traditional image retrieval is a kind of text retrieval in essence. Every image in the image database needs to be induced and noted ahead of time by manpower. The image retrieval depends totally on the names of images and notes which make the workload heavy and notes can't update automatically. Anymore, the limited method of describing images and serious subjectivity make the retrieval result wrong.In order to solve the problem mentioned above, we not merely need to extract images features automatically in an all-round way objectively, and we will also need to work out techniques of content-based image retrieval. Since initial stage of 1990s, the technology of utilizing the content of images, such as color, shape and texture etc. has arisen, which is named content-based image retrieval. Content-based image retrieval obtains the content of images according to the search, analysis and understanding of visual media from lower to high level. It has important research meaning and enormous practical value, so it becomes study focus rapidly. There are many systems shaped both at home and abroad at present, such as QBIC of IBM, Photobook of MIT, VisualSEEK of Columbia etc. Though techniques of content-based image retrieval have a large amount of studying and real application, a lot of technology including unripe, also need further research and improvement. Now there is no system suitable for any kind of image retrieval field at present, and it is a very good identification that search of pictures of large-scale retrieval system such as present Baidu and Google. So the research on techniques of content-based image retrieval is still having a long way to go.Color is an attribute of object's surface, and is a kind of visual feature. Each object has itsown unique color attributes. The color is a basic element of image content. Color feature is. more relatively stable than other features. It is invariant to rotation, translation, displacement,and scaling, as well as some other deformations and it shows very strong robustness. Moreover,color feature becomes one of the most widely used features in the existing retrieval system because of its simple calculation and convenient processing. Just because of these characteristics of the color, it is a most extensive and widely used feature in the retrieval system. So the vast majority of content-based image retrieval systems use it as an image essential feature. Because of the importance of color and it's shortcoming in color feature extraction and matching, we will still discuss and study it. In this paper, we analyze various color spaces and put forward a proposed non-uniform quantization model and a fuzzy color quantization method.Texture can be regarded as the repeated duplication of certain elements or regular permutation association of some lamination element. These characteristics have more obvious embodiment in the pictures such as cloth, ripples, building materials images, etc.. The texture feature is more suitable to search above kind of pictures. But the images which have no texture or little texture couldn't use the feature to retrieval. Therefore we seldom use the texture feature in the image retrieval system. We often use texture feature to retrieve images with other image features. Because of the importance in the fields such as pattern-recognition and computer vision etc., the texture analysis has been an important research direction in computer vision field and has achieved great achievement over the past three decades.Shape features play an important role in the human visual perception, awareness and understanding, therefore shape features have been a kind of important feature in the field of image expression and image retrieval. Shape feature describes the high-level visual features (such as target, object, etc.), therefore it is important in the acquisition of image semantics. But unlike the low-level features such as color or texture, the expression of shape features must be based on image segmentation. However, the current shape description method is not very mature, so it is difficult to achieve accuracy, fast automatic image segmentation and shape feature extraction, and image recognition or retrieval will lead to application limitation. The shape feature is only used in certain specific applications, or needs some professional knowledge in the field to improve the accuracy of shape recognition, or retrieval, such as face recognition, iris recognition and fingerprint identification system. Moreover, people are not sensitive to the shape of the object, rotation and scaling, so the appropriate features should meet the need for transformation, rotation and scaling, which brings difficulties in the calculation of the shape similarity. Therefore, it is limited in some professional library of applications, such as trademarks, medical images and biological images.The significance of image segmentation is to divide the image into meaningful and disjoint regions which have similar characteristics, and then extract the various features of these regions for image analysis, understanding and image retrieval. Therefore, research on image segmentation in the field of image retrieval is of great significance and it has been a research hotspot in the field of image processing. There is no. universal image segmentation theory so far, but there have been a large number of new theories, other disciplines and new methods of combining research results. The existing image segmentation methods, despite many species, however, have some limitations, and applications are often split in a good effect, while the effect will be less than satisfactory in the other areas of image segmentation. This is because each of the segmentations of the image has its own features, so automatic segmentation method is very difficult to apply automatically and accurately in any field segmentation. Thus, segmentation technology is a hot focus and difficulty on research, which requires that the researchers continue their efforts. An image segmentation based on graph theory approach is put forward, and is applied to the brain medical image data mining system, thereby the classification of medical images of the brain has been finished.With the development of network technology and multimedia technology, content-based image retrieval technology becomes a research hotspot. However, there is a semantic gap between visual low-level image features and high-level image features, which leads to unsatisfactory retrieval results. To solve this problem, the introduction of human-computer interaction image retrieval relevance feedback technology effectively improves the retrieval performance of the system. Relevance feedback systems and human interaction is the process of the system to make the computer to understand the image information, the system simulation of human's observation and understanding of the images so as to enhance retrieval performance. Although relevance feedback is essentially based on the low-level physical characteristics, but the semantics are user-oriented. Different users have different needs, and how to make the system grasp accurately the users' needs, the relevance feedback is a better way. This paper introduces the relevance feedback technology, basic knowledge and applied background, proposes particle swarm optimization algorithm (Particle Swarm Optimization, referred to as PSO) and the introduction of content-based image retrieval relevance feedback, and lets the system get human's understanding of the image information, simulation of human observation and understanding of images so as to enhance image retrieval performance.Relevance feedback is an online system which needs high real-time. Therefore, the system running speed is very important. Feature dimension and feedback algorithm make a tremendous influence on calculating time. So, it is our work in the future to reduce the feature dimension of the picture and improve the feedback algorithms. Furthermore, the feedback system should be given to the user evaluation and accumulation of image-related information storage in order to be used again in the future. Studying the long-term learning strategy, setting up a feedback retrieval of memory, reducing the time of follow-up image retrieval and searching space and enhancing the retrieval efficiency of overall image retrieval system is our future research work.Besides image feature extraction, image segmentation, image feedback technology, we have also finished two retrieval systems. First, the Web-based image retrieval system is finished, which uses PSO algorithm to adjust dynamically the weights of the various features so as to enhance retrieval precision and speed. The second system is finished, and it is based on Weka, and uses Weka's data pre-processing, classification and visualization to realize the system based on feedback, which lays the foundation for further study.