Robust Image Classification with Context and Rejection
2018-12-14T17:53:58Z (GMT) by
Classifications systems are ubiquitous; despite efforts going into training and<br>feature selection, misclassifications occur and their effects can be critical. This is<br>particularly true in classification problems where overlapping classes, small or incomplete<br>training sets, and unknown classes occur. In this thesis, we mitigate misclassifications<br>and their effects by adapting the behavior of the classifier on samples<br>with high potential for misclassification through the use of robust classification<br>schemes that combine context and rejection. We thus combine the advantages of<br>using contextual priors in classification with those of classification with rejection. In<br>classification with rejection, we are able to improve classification performance at the<br>expense of not classifying the entire data set.<br>We thus add the following tools to the robust classification toolbox: 1) we derive<br>performance measures for evaluating of classifiers with rejection; 2) we create<br>a family of convex algorithms, SegSALSA, to classify with context; 3) we design<br>architectures for robust classification with context and rejection that encompass interactions<br>between context and rejection. We validate our approach on two different<br>real-world data sets: histopathological and hyperspectral images.