Developing machine learning techniques for real world applications

Machine learning techniques have been successfully applied to computer vision area recently. In this thesis, we have proposed two novel machine learning techniques, a novel semi-supervised learning framework and a reliable graph model, and have applied these two techniques to three real world computer vision applications successfully. Semi-supervised learning (SSL) has become an important tool for learning the classifiers which only have limited labelled training samples, together with a large number of unlabelled training samples. Typically, the semi-supervised learning is achieved through iterative supervised learning. However, the existing stop criteria of semi-supervised learning algorithms are ad hoc and they cannot guarantee that the accuracy increases when the number of iterations increases. Besides, current SSL algorithms do not consider the situation that different applications may have different accuracy requirements on different classes. In this thesis, we present a novel framework which contains an Accuracy Probabilistically Increased stop Criterion (APIC) and an Application Oriented Labelling strategy (AOL). We prove in theory that a semi-supervised learning algorithm under this framework is probabilistically guaranteed to have the accuracy increased when the number of iterations increases. We also prove that the application oriented accuracy is maximized at each iteration of the learning procedure. This framework is called Enhanced Semi-supervised Learning (ESL). A semisupervised learning algorithm under the ESL framework is named as an ESL algorithm. Extensive evaluations of the ESL framework based on two well-known semi-supervised learning algorithms using the public data sets from the UCI Machine Learning Repository clearly indicate the superiority of the ESL algorithms to the original semi-supervised learning algorithms. In order to further evaluate the ESL framework, we have applied it to two real applications. The first application is object detection in aerial imagery, which has been well studied in computer vision for years. However, given the complexity of large variations of the appearance of the object and the background in a typical aerial image, a robust and efficient detection is still considered as an open and challenging problem. In this thesis, based on the ESL framework, we have proposed a novel context-based detection methodology. The performance evaluations of this detection methodology against a state-of-the-art object detection algorithm and another detection algorithm without using the ESL framework clearly demonstrate the promise and the superiority of this approach, as well as its contributions to the advancement of the computer vision literature. The second application is automatic medical image retrieval. The demand for automatically annotating and retrieving medical images is growing faster than ever. In this thesis, we present a novel medical image retrieval method for a special medical image retrieval problem where the images in the retrieval database can be annotated into one of the pre-defined labels. Even more, a user may query the database with an image that is close to but not exactly what he/she expects. The retrieval consists of the deducible retrieval and the traditional retrieval. The deducible retrieval is a special semantic retrieval and is to retrieve the label that a user expects while the traditional retrieval is to retrieve the images in the database which belong to this label and are most similar to the query image in appearance. The deducible retrieval is achieved using SEMI-supervised Semantic Error-Correcting output Codes (SEMI-SECC). The active learning method is also exploited to further reduce the number of the required ground truthed training images. Relevance feedbacks (RFs) are used in both retrieval steps: in the deducible retrieval, RF acts as a short-term memory feedback and helps identify the label that a user expects; in the traditional retrieval, RF acts as a long-term memory feedback and helps ground truth the unlabelled training images in the database. The experimental results on IMAGECLEF 2005 annotation data set clearly show the strength and the promise of the presented methods. A hierarchical shadow detection algorithm based on a novel proposed machine learning methodology, i.e., the reliable graph, for color aerial images is also presented in this thesis in order to meet the two challenges for static shadow detection in the literature: different brightness and illumination conditions in different images and the complexity of aerial images. The hierarchical algorithm consists of two levels of processing: the pixel level classification, achieved through modelling an image as a reliable graph (RG) and maximizing the graph reliability using the EM algorithm, and the region level verification, achieved through minimizing the Bayesian error by further exploiting the domain knowledge. Further analyses show that Markov Random Field model based segmentation is a special case of the RG model. The relationship between the RG model and the relaxation labelling (RL) model is also discussed. A quantitative comparison between this method and a state-of-the-art shadow detection algorithm clearly indicates that this method is promising for delivering effective shadow detection performance under different illumination and brightness conditions.