Robust Image Classification with Context and Rejection

Classifications systems are ubiquitous; despite efforts going into training and
feature selection, misclassifications occur and their effects can be critical. This is
particularly true in classification problems where overlapping classes, small or incomplete
training sets, and unknown classes occur. In this thesis, we mitigate misclassifications
and their effects by adapting the behavior of the classifier on samples
with high potential for misclassification through the use of robust classification
schemes that combine context and rejection. We thus combine the advantages of
using contextual priors in classification with those of classification with rejection. In
classification with rejection, we are able to improve classification performance at the
expense of not classifying the entire data set.
We thus add the following tools to the robust classification toolbox: 1) we derive
performance measures for evaluating of classifiers with rejection; 2) we create
a family of convex algorithms, SegSALSA, to classify with context; 3) we design
architectures for robust classification with context and rejection that encompass interactions
between context and rejection. We validate our approach on two different
real-world data sets: histopathological and hyperspectral images.