Research Of Image Classification Based On Multi-instance Learning

Multi-instance learning is proposed as a new kind of machine learning framework based on complex practical problems. As an effective way to deal with multiple semantic features, it can be applied into image classification. Multi-instance multi-label learning is regarded as the combination of multi-instance learning and multi-label learning. Compared with multi-instance learning and multi-label learning, multi-instance multi-label learning can describe objects more comprehensively, and it provides a new train of thought for multi-class classification, so it has important scientific significance and practical application value for image classification and recognition. In this paper, we study the application of multi-instance multi-label learning in image classification, as well as text categorization, audio classification.(1) We introduce the related theory and model of support vector machines(SVM), analyze the multi-instance learning framework and three kinds of classic multi-instance learning algorithms, the concept and the framework of multi-instance multi-label learning. Then we study two classical classification algorithms based on multi-instance multi-label learning.(2) We analyze and study the method of representing instance correlations in a bag in multi-instance learning, as well as the multi-kernel learning algorithm. By introducing the instance correlation features into the multi-instance multi-label learning, and introducing the multi-kernel into constructing classifiers, we propose a multi-instance multi-label learning algorithm based on multi-kernel fusion. Simulation experiments on scene image data sets, text data sets and audio data sets have verified the effectiveness of this algorithm on dealing with classification problems.(3) We study the instance selection method via joint 2,1l-norms constraint, and introduce it into the multi-instance multi-label learning to eliminate instances with interference and select representative instances. Considering the label correlations, we construct the classifier based on label correlations and propose a multi-instance multi-label leaning algorithm based on feature selection.