Research On A Kind Of Image Retrieval Framework And Implementation Of Its Prototype

With the expansion of a large number of images, it become a very important issue in the field of image's application that how to retrieve the image wanted efficiently and quickly. So content-based image retrieval (CBIR) is now the study hot-point in the world. It extract low-level visual features such as color, texture, shape etc, but these low-level features contain little on high-level semantic contents. And traditional text annotation is simple and clear on expressing high-level semantics. In this paper, we first discuss the image retrieval methods based on low-level visual features, then discuss a active learning framework combining these two methods.The active learning framework is proposed recently by American scholars. In this paper we analyze and improve the framework. The framework suppose that the image's semantic feature can be expressed by using a multi-level attribute tree, and the probability of a certain image having a certain attribute can be estimated by interpolation method using the neighboring image's value. Here, the neighboring relation refers to the neighboring in the low-level feature space. The framework direct the annotation process using the principle of max knowledge gain, not random annotation method, in order to gain better retrieval result in the same number of annotated images.In this paper we use the color auto-correlogram as the similarity metrics of images in low-level feature space, and change the bandwidth function. Then we propose the semantic relevance feedback. The system react differently to the positive and negative user's feedback so that the system can go on learning after the annotation process by updating the probabilities of the list of attributes of the relevant images and reaching the real values. Last, we design and implement the experimental system using C++Builder6. The experimental system extract the low-level features of images such as HSV histogram, the texture got from coexistence matrix, color correlogram, and according to the characteristic of our image database, design the evaluation function such as the average rank ratio to evaluate and compare the performance of different integration of different features including semantic, and validate the active effect of feedback using experiment results.