Multi-Points Diverse Density Algorithm And Its Application In Image Retrieval

The twenty-one century is an information explosion century. The development of the multimedia, computer and communication technology is the direct result of more and more information into people's lives. Image is one of the most important means of carrying information. How to organize, run and utilize image is very important, which make CBIR be one of the most active research focus in multimedia. The effective extraction and matching of image features are the key technologies in CBIR, on which most researchers concentrate.In this paper, we studied the CBIR and discussed the extraction algorithms of image feature. We see that traditional CBIR use the low lever and global feature to describe the image, obviously, it is not enough and perfect that the image content is described by the single and global feature, not to mention the interesting section of the user. This point makes a great influence on the image retrieval.The Multiple Instance Learning is a new frame in the Machine Learning Word. The multiple-instance, as the name suggests, uses several instance to describe the image, which just deals with the problem mentioned in last part well, so the researchers in this filed introduce MIL into the CBIR, and they have gotten a large achievement. There are two steps when apply the MIL: firstly, you must partition the image into several blocks and extract the low lever feature of each block, which are made of the instance bag. Secondly, this paper uses MIL algorithms to extract the key information of the user wanted, which is the goal feature when we process the image retrieval.In this paper, the concept and the algorithms of MIL are studied in detail, which are proposed in recent years. By the studying, it is found that all the MIL algorithms described the image content with the single instance point most times, including the classic Diverse Density algorithm and Expectation-Maximization version of Diverse Density. It is one-sided that describing the image contents, which include rich and diverse information, with the single point. Address this problem, the Multiple Point Divers Density algorithm is proposed based on DD. MPDD regards all the density points as the goal features describing the image content, which overcomes the problem of DD that uses the single density point to describe the image.To applying MIL into CBIR, the sample image is partitioned into 36 uniform blocks firstly, secondly, the feature of color histogram of 72 dimensions and co-occurrence matrix 10 dimensions are extracted. As a result, a sample image can be described with 36 feature vectors each of which has 82 dimensions.The way of uniform block is simple and easy, but it is not reasonable, so the K-Means clustering algorithm is applied when the image is partitioned into blocks. The clustering algorithm is unsupervised. The criteria of K-Means clustering algorithm is that the similar but not identical data will be classified into a same class, otherwise, a different class. The difference between classes will be larger, but in a class, the difference will be smaller. By parting the image and extracting the features, and then classifying them, the regions with same features will be classified same clustering and labeled the same instance; the ones with different features will be classified into different clustering and labeled the different instance. This process is regarded as the second partition, which makes the concept consistent and integrity and reduces the computation and memory at the same time. The number of clustering is be set 5 which got by the peak numbers of the color histogram.After clustering, the sample image is expressed as a bag of 5 features of 82D. When the bag is ready, the MPDD algorithm can be used to learning it. Because the single point is not enough to express the image content, MPDD outputs all diverse points which are gotten by gradient descent algorithm, and expresses the image bag. This way which makes the image complete and comprehensive plays a significant role in improving the efficiency of image retrieval.At last, experiments are made on the bases of the DD and MPDD algorithms, the images of which are belonged the Corel database, in which has 1000 images total. By the many experiments, this paper makes analysis and comparison thoroughly between DD and MPDD by calculating the Recall and Precision, drawing the figure of"averrec-averpre"and computing the average time of retrieve. The results show that the MPDD algorithms proposed in this paper is more effective than DD, which plays a positive role in image retrieval and makes a solid foundation for further research.