Machine Learning for Image Classification

Machine learning is a powerful tool for pattern classification. The goal of machine learning is to develop methods that can automatically detect patterns in data, and then to use the uncovered patterns to predict future data or other outcomes of interest. Machine learning is thus closely related to the fields of statistics and data mining, but differs slightly in terms of its emphasis and terminology (Murphy, 2012). Machine learning uses the theory of statistics in building mathematical models, because the core task is making inference from a sample. A model may be predictive to make predictions in the feature, or descriptive to gain knowledge from data, or both (Alpaydin, 2010). Recent year, image classification has become quite a significant topic in image engineering (Zhang, 2013a) with the requirements from action recognition (Zheng et al., 2012), emotional image retrieval (Li & Zhang, 2010), scene categorization (Liu & Zhang, 2011) and behavior understanding (Zhang, 2013b), etc. Image classification aims at associating different images with some semantic labels to represent the image contents abstractly. To achieve this goal, various machine learning and pattern recognition techniques could be used (Bishop, 2006). Among many potential techniques adopted in image classification, the technique that uses dictionary learned by sparse coding has achieved competitive performance recently. Sparse coding is capable of reducing the reconstruction error in transforming low-level descriptors into compact mid-level features. However, dictionary learned only by sparse coding does not have the ability to distinguish different classes, and it is not the optimum dictionary for the classification task. In this article, some current techniques for image classification and their existing problems are reviewed, some research directions are discussed, and some progress works, especially a novel discriminant dictionary learning method combining linear discriminant analysis with sparse coding and an integrated dictionary learning technique on manifold are introduced. Their performances are compared with other techniques and quite satisfactory results have been obtained.