Statistical classification of images

Resum

Image classification is a burgeoning field of study. Despite the advances achieved in this camp, there is no general agreement about what is the most effective methods for the classification of digital images. This dissertation contributes to this line of research by developing different statistical methods aim to classifying digital images. In Chapter 1 we introduce basic concepts of image classification and review some results and methodologies proposed previously in the literature. In Chapter 2 we propose a method to classify images by their content. We are able to distinguish between landscape from non-landscape pictures by using three features obtained directly from images. We obtain better classification rates than those obtained by other authors dealing with similar kind of scene classification. In Chapter 3 we address the handwritten digit recognition. We suggest a set of intuitive features to perform the classification. Since the features are calculated with the binary image, we propose a novel technique to obtain the optimum threshold to binarize images, based on statistical concepts associated to the written trace of the digit. The classification is conducted by applying multivariate and probabilistic approaches, concluding that both methods provide similar results in terms of test-error rate (3.5\%). In Chapter 4 we propose the application of Functional Data Analysis to analyze and classify images. While a limited number of authors have suggested the application of FDA for image classification [\cite{Batis10}], we suggest that this branch of statistics has represents a promising approach and offers several avenues for future research. We close the dissertation in Chapter 5 with a set of concluding remarks. Overall, the methods suggested in this dissertation are simple to apply, intuitive in their interpretation and their performance is comparable with other complex methods applied to the same problem. Moreover, the features suggested require less processing time than other methods (as support vector machine classifiers) and therefore require less computational capacity.