Image classification is an important application of machine learning. It can provide supports to many other vision tasks, such as object detection, CBIR etc., as well as meet real world practical needs. Generally speaking, the generalization ability of image classification is not only related to the classifiers, but also dependent on the features used. With the development of Internet and mobile devices, more and more images are collected. In the meantime, the device which carries image classification system has gradually shifted from personal computers to mobile devices on which traditional approaches cannot adapt. As a result, how to automatically classify large-scale image data effectively and efficiently has become an important research topic.Currently image features include color, texture, shape and key points, etc. Most systems directly merge these features in a whole feature vector or in a weighted style, and these approaches may not make full use of the information provided by the feature. In the field of semi-supervised learning, a lot of researchers interest in multi-view learning, Multi-view co-training can make use of unlabeled data to enhance generalization of classifiers. Inspired by this, if we treat different kinds of feature as multiple views of image, then multi-view learning can well integrate information provided by each view. In particular, on real world applications with real-time requirements, we can only extract a few parts of features and in this case, with multi-view learning we can choose the best features needed. Besides, multi-view learning can also use multiple features to help each other, so that the discriminant ability of a single feature can be enhanced.For image classification on huge data, classical machine learning algorithms which based on single machine cannot adapt to the explosive growth of data. So, there is a need to port image classification algorithms to distributed computing framework. How to implement the algorithm efficiently on distributed environments is also a problem which needs to be investigated.This article will present solutions for large-scale image classification with the aspect of multi-view learning, efficient feature extraction and distributed modeling.In chapter 2, we propose Auxiliary Modal Classification (AMC) approach. AMC incorporates the informative strong features to help the weak features training by minimizing the prediction gap of unlabeled data between these two models. In the training phase, both strong and weak features are trained to be consistent, i.e., the part of weak features is adjusted to have a similar prediction as the strong part on a large amount of unlabeled data. In the test phase, however, only the features from weak view, which are generally with low extraction cost, are needed, so that AMC can work efficiently and effectively.In chapter 3, we propose Discriminative Feature Extraction (DFE) algorithm. DFE extracts features in a sequential manner. It sorts each feature in a descent order by the discriminant ability on each particular instance, so that feature with higher discriminant power can be extracted first. It is notable that DFE identifies the unique contribution of each feature instance specifically, hence the extracting sequence is different among data, and as a sequence it can dramatically reduce the feature extraction costs.In chapter 4, we propose Large Scale Support Vector Machine (LS2VM) algorithm. LS2VM is a parallelized SVM by separating multi-class problem into several independent binary problems. In addition, we cached support vectors in LS2VM to meet memory limit of each map-reduce node and training the support vector machine in an iterative style. Experiments on real datasets show that LS2VM has a good efficiency and scalability while dealing with big data. |