Active And Semi-Supervised Learning Based On Elm For Multi-Class Image Classification

Image classification is an image processing method of making a distinction among different categories according to different characteristics. The traditional supervised classification algorithms argue that all the training instances should be labeled before the classifier has been built. Nowadays, the available instances become more and more. It will cost much resource and time by labeling all of them for supervised algorithms.Therefore, we propose algorithms combining the active learning and semi-supervised learning to improve the classification performance with unlabeled instances. Considering the real-time modeling for big data, we choose ELM which based on the SLFNs as the final classifier instead of SVM. Compared with the traditional neural networks, ELM is not only more simple, but also faster because of randomly choosing hidden nodes and generating the input weights and hidden bias, thus ELM is suitable for the multi-class image classification.The thesis mainly focuses on the study of multi-class image classification combining the active and semi-supervised learning based on ELM, which includes the following contents:(1) Study on the image pre-processing and classifier modeling. We take advantage of the Bag of Words(BOW) model for the pre-processing. Firstly, we extract SIFT features from the raw images. Then we achieve.K clusters (K vocabularies) with K-means algorithm for all the features. Finally, we get the word frequency statistics histogram by feature coding as image representation, and take them as inputs of ELM classifier.(2) Study on the multi-class image classification by combining the BVSB active learning and the distance-based self-training semi-supervised learning algorithm based on ELM. Firstly, considering that both active learning and semi-supervised learning algorithm must measure the uncertainty of unlabeled instances, we transform the actual output of in ELM into the approximate posterior probability. Then we propose the algorithm combining BVSB and the nearest-neighbor self-training(NN-ST) based on ELM, which guarantees to improve the quality of the classification model with labeling instances as few as possible.(3) Study on the multi-class image classification by combining the BVSB and manifold regularization theory-based global semi-supervised ELM algorithm. Considering that the proposed semi-supervised learning algorithm above only concentrates on the local information of the unlabeled instances, we try to use the global information of them in the semi-supervised learning, so that the classification quality can be further improved.