| Image retrieval based on multi-instance learning was one of the most prominentareas in pattern recognition in recent ten years. Numerous MIL image retrieval methodshave been proposed.Current systems can do fairly accurate recognition underconstrained scene using these MIL methods. It has attracted much attention due to itspotential application values. Nowadays, Image retrieval based on multi-instancelearning has great value in image preprocessing, regional image annotation, standardimage retrieval, face recognition and so on.Requirements of the practicality of the Image retrieval based on MIL are increasedalong with the wide use. Also, most of the methods can only do fairly accuraterecognition in specific environment. But the specific environment is not suitable forother applications. This paper do an in-depth study on the problems mentioned above,The main achievements are as follows：1. To improve the image retrieval precision effectively by using tags influencefactorThe same picture has different contribution on the central idea in different tags. Soit’s important to choose a suitable influence factor about tag in setting up a math modelof html document. Based on some new concepts, such as ttf (Term Frequency in Tag), itf(Inverse Tag Frequency), a document can be transformed to a matrix which rowrepresent html tags and column represent keywords. Then let’s calculate the averagedistance between every vector to centroid in the list which be constituted of certain rowin every document. By the average distance we can get the tag’s influence factor. We putthis assistive technology applied into VSM algorithm to build association betweendemand and target. Experimental results show that when tag’s influence factor beapplied in the new documents, the results is better than before.2. The keyword is labeled effectively in image segmentation region bymulti-instance learning methodIn most existing training data set for image annotation, keywords are usuallyassociated with images instead of regions. Because not use the information in regions, itis difficult to use multi instance learning methods for image annotation. In addition, oneto many relationship between keyword an region is important to the high efficientsearching, so it is an urgent demand for elimination of synonyms. So a improved MILarithmetic(AFSVM-MIL) are bring forward, which based on active learning fuzzy support vector machine. This method classify the training datasets by using thesimilarity between individual images, then, keywords are labeled on the instanceseffectively and the labor intensity of manual annotation are reduced. Experimentalresults show that algorithm is not only adaptive but also accurate. Even to the extentthat, keywords can be labeled on instances one by one.3. Do in-depth research and improvement on image retrieval based on bothmultiple instance learning and hausdorff algorithmIn the process of multi-instance learning, cost of each instance shows thecorrelativity between instance and searcher’s requirement. The past Hausdorff’sdistance arithmetic think nothing about cost of each instance. So the result can not meetto searcher’s demand. And the past Hausdorff’s distance arithmetic either is sensitive toisolated point or just think about the nearest twe instances. Thus a improved Hausdorff′sDistance arithmetic is presented which adjust distance by cost of instance and overcomethe twe shortcoming above. It is used in k-means arithmetic to solve the semantic imageretrival problem. Experimental results show that this algorithm is feasible and theperformance is superior to the other k-means algorithms.4. Putting forward the EMD-MIL framework based on information fusiontechnique and rapid technology, so application of multi-instance learning in facerecognition run wellA number of factors, such as brightness, shooting angle and the lack considerationof entirety, have a negative effect on many face recognition algorithms which werebased on the facial features. And earth mover’s distance, as a distance which was fit formulti-instance learning, run slowly on large data sets. Focusing on the two problemsabove, a fast face recognition algorithm is proposed. The algorithm is based onEMD-MIL framework and information fusion technique. Firstly, both informationfusion MIL framework and distance threshold were used to to make algorithm performwell. Secondly, relying on instances which be composed of the three facial features andthe overall character, a fast EMD-MIL framework was bring out. The fast EMD-MILframework run more quickly than the old edition. Experimental results on the ORL andthe MIT show that this algorithm is feasible and the performance is superior to someother similar algorithms.5. It is difficult to apply traditional multiple instance learning in face recognition.A method called instance equity can resolve this problem. This method give a way toget local and total equity and make the face recognition runs more effectively The traditional voting mechanism in Multiple instance learning is prone tomisunderstanding, because the local similarity does not mean the overall similarity inface recognition based on facial features. While the traditional selection method for theratio of fusion need to be optimized urgently. Therefore the traditional MIL-Frameworkwas improved. Firstly, instance equity concepts and its calculation method werepresented according to the sample characteristic.Each kind of instance has differentequity. So packet’s classification probability was derived from the sum of multiplicationattribute and its equity. Secondly, the overall characteristic be considered as a specialinstance, and the overall sample equity threshold be used to control equity ratio. At thesame time, abnormal point generation rate be reduced by means of the feature fusion.And the recognition rate improves with the help of threshold optimization selection.Experimental results on the ORL and MIT image set show that this algorithm is feasible,and the performance is superior to other MIL algorithms. |
PDF Full Text Request