Research On Image Retrieval Based On The Multiple-instance Learning

Vision is the most important way to obtain information. Pictures are usually able to express more than the text, sound. In recent years, with the popularity of image acquisition equipment there are massive image data generation every day. Therefore, it is valuable and meaningful to obtain the image of interest in real time. The general semantic image retrieval system which based on machine learning is trained by extracting the features overall the total image. As we know, the images from the real world often contains multiple objects. So the image level feature is usually contaminated by other objects, resulting in a not very good training model. In this paper, a model of image automatic semantic annotation based on multi instance learning is proposed, which can regarded the image as a bag containing multiple instance or objects and learn the feature of the object in interest. The method can effectively solve the problem that the machine learning method is difficult to overcome. It should be noted that these semantics are addressed here mainly referring to the proper nouns and categories nouns.In order to be able to divide the image into meaningful regions quickly and effectively, a fast image segmentation algorithm based on graph is used in our subject. By defining and testing the method of calculating the similarity of different pixels to select the appropriate similarity measure and the evaluation criteria and the experimental parameters are defined to achieve the expected results. In order to extract discriminative feature, this paper uses a deep convolutional network model which achieved great success in image classification. A network model with the ability to extract the discriminative features is obtained by fine tuning of the “Googlenet” network. And the network implementation is achieved by defining the recognition and verification signal object function. And then, based on the study of multi instance learning algorithm, we find that the feature of the negative examples is the key to guide the learning algorithm to eliminate the non-interested object features and convergence to the true features. So we need to choose the training sample with the target that the negative bag can contains the non-interested objects which contains in the positive package for training the multiple instance learning model. In order to use all the classification models to get a picture annotation model, we use them in the form of “vote”.In the research, we use binary multiple instance classifier. And training the model for each sematic. In the experiment,we choose twenty eight categories from the Core Image Dataset. And there are nearly twenty eight hundreds pictures in the data. The mean test accuracy of each model is 88.48%. And the semantic annotation model is defined by using all the classification models in the form of “vote”, and the test accuracy of the data set which is mixed in all classes is reached 86.25%. And finally we set up a simple semantic image retrieval system based on image semantic annotation model proposed by our research.