Natural Scene Image Classification Based On Multi-instance Multi-label Learning

In recent years, with the rapid development of internet and the widespread use of digital products such as digital cameras, computers and smart phones, the image information has been increased explosively. That how to classify these images has become a problem to be solved. For complex images in real life such as natural scene image, the traditional supervised learning framework does not meet the requirements. So the scholars have put forward the multiple instances learning, multiple labels learning and multi-instance multi-label(MIML) learning framework.This paper mainly applies the multi-instance multi-label learning framework to natural scene image classification. Study and learn some multi-instance multi-label learning algorithms and improve the MIML algorithm based on RBF neural network. In this paper, the main works are described as follows:(1) The MIML learning algorithms process the data bags that contain more than one instance, the performance of the bag generation methods will directly affect the final classification results. Study and implement two kinds of bag generation methods that based on fixed area and image segmentation. The experiment shows that the SBN algorithm has a better result for natural scene image classification.(2) Study and implement the MIMLBOOST, MIMLSVM, M3 MIML and MIML-KNN algorithm for natural scene image classification: MIMLBOOST and MIMLSVM are based on the degeneration and take the hypothesis that the instances or the labels are independent. But they lose a lot of useful information during the reduction process. M3 MIML and MIML-KNN algorithm consider the correlation between the instances and the labels. The experiment shows that M3 MIML and MIML-KNN have better results than MIMLBOOST and MIMLSVM on image classification.(3) Do further study to MIML based on neural network and mainly analyze the MIML algorithm that based on RBF neural network. Then improve the algorithm based on RBF neural network---- introducing an adaptive adjustment coefficient to the average Hausdorff distance. The experiment shows that: ? The algorithm based on RBF network has a better classification result than the algorithm based on BP network and other traditional algorithms for natural scene image classification. ?Analysis four different Hausdorff distance: maximum Hausdorff, minimum Hausdorff, average Hausdorff and their average value. The experiment shows that the algorithm using the average Hausdorff distance has the best result. ?For the problem that the average Hausdorff distance weakens the role of the minimum distance between the instances in the two bags, the modified average Haudorff distance can further improve the accuracy of classification.