A Multi-instance Image Retrieval Method Based On Key Points Mapping And The Total Probability Model

Great changes have been taken place in computer science since we entered the 21 st century. That people do research on computer version becomes more and more widely. Image retrieval, as a hot topic in the field of computer vision research also attracts many scholars. It becomes an urgent need that people choose the image which they are interesting in from the rapid growth of database. Because many images always have many semantic sections and that is consistent with MIL, the problem of image retrieval becomes a very challenging topic. The disadvantage of the former is depending on the work of a lot manual text annotation, and the latter’s accuracy is limited by the image’s semantics in spite of using the color and the texture and the shape features.Multiple instance learning (MIL) is a new method of the machine learning methods. It is believed that the training set is composed of a set of labeled bags, and each bag contains a set of instances, but the label of an instance is not certain. The positive bag contains one positive instance at least, and the instance’s label is negative when it belongs to a negative bag. This description coincides with the characteristics of the image’s multiple semantics. So it is feasible to solve the semantic problem of image retrieval by using MIL. Now many scholars have applied it to the field of image retrieval, they try to improve the algorithm of MIL image retrieval by applying different method of the bag generator and the MIL methods. The methods of classic bag generator are SBN (single blob with neighbors) and RR (Reasoning with Regions) and so on. The methods of classic MIL are the algorithm of EM-DD and MILES and so on.Among the methods of the bag generator, the method of SBN has the characteristics of easy extraction and fast speed, but its accuracy is general, and the method of RR is more accurate than the former, but its extraction is more complex. Both of them have not considered the region of interest from the image itself. The paper uses the key points and the super-pixel regions to generate the bag so that the bag has the characteristic of interest. Among the methods of MIL, the EM-DD method need a lot time to calculate the diverse density points. However, the MILES method gets the bag from the feature prototype space of the training set, and then it converts the MIL problem into standard supervised learning problems. The MILES method has a more mature theory and more simple realization characteristics by contrast. However, because there are many instances in the feature prototype space of the training set and the noises of them reduce the MILES’s efficiency and accuracy. This paper also uses MILES’s core idea, but this paper uses the visual semantic model (VSM) to solve the problem of the MILES method, and this paper also uses the Total Probability Model (TPM) to improve the MILES according to the shortcomings of not making full use of the correlation between the instances and the important degree of instances in the image. It is proved that the correlation based on TPM is a general method to calculate the value of correlation between the instance and the clustering center.A multi-instance image retrieval method based on key points mapping and the TPM is proposed in this paper. It regards an image as a bag and the regions in the image as instances. Firstly, each image is segmented by the method of Simple Linear Iterative Clustering (SLIC). Secondly, the key points are found out from the image and map the points to the regions. Thirdly, the features of color and texture are extracted from those super-pixels which are mapped by the key points, and the bag is generated by the feature space. Finally, the feature of the bag will be produced by using the VSM and TPM, and we use a method of Manifold Ranking to calculate the saliency values of those super-pixels which reduce the influence of the super-pixels which have low saliency values but include key points. Lastly, the Support Vector Machine is used to build the query model and the query results are sorted according to the correlation.The advantages of this paper’s method are as follows. Firstly, the multi-instance bag is generated from the region where the users are most likely interested in, which means that the bag makes full use of the information from the regions which place the key points belongs to. Secondly, this paper use VSM to solve the problem that there are many instances in the feature prototype space of the training set and they may cause noises. Lastly, this paper uses the TPM to extract features of the bags, and both of the correlation between the instances and the important degree of instances in the image are fully taken into account. It is proved that our method has better accuracy by comparing with other methods through the experiments on the image datasets of the Corel and SIVAL benchmarks.