A comparative study of image invariants for text / non-text classification



Image recognition main objective is to enable computers to recognize shapes without needing any human intervention. However, one problem in this field is automatic recognition regardless to image transformations like rotation or scales changes. Image invariants based on moments have received a particular attention since they respect fully or partially the constraints listed above. Various shape-based invariant algorithms are popular in Document Image Analysis and especially for Optical Character Recognition systems since they provide relevant features used to differentiate characters. Nevertheless, they also find an interesting application in Document Layout Analysis by providing information that can be used to distinguish text from non-text elements. Thus, we will introduce the concepts of image invariants and evaluate the performances of six state-of-the-art shape-based image invariants for text / non-text classification. An attempt of an image invariant algorithm inspired from the principle of compressive sensing is also offered.