On the Definition of Categories for Image Classification Evaluation

This paper will study the process of creation of categories for the evaluation of algorithms for semantic classification. In many papers on the topic, one can see that algorithms and methods are evaluated over a set of often haphazardly defined classes, with little or no coherence between them. We begin with a brief reminder of the process of semantic formation in images, and show that, because of the heavy cultural involvements in the process, it is unfeasible to create “semantic” categories for algorithms evaluation. We find a solution in a concept from structural semantics: the semiological level of evaluation, which defines the core of a sign, that is, the portion of meaning that does not depend on context. We analyze the properties of the immanent semantic space, its semantic axes, and the concept of seme. Armed with these instruments, we outline a process for the creation of a coherent set of categories for image analysis evaluation.