Method for detecting false responses of localization and identification algorithms using global features

The paper presents a method for detecting false responses of localization and identification algorithms. The method considers matching image characteristics that cannot be described by local features stably and completely. It is proposed to use image zones containing such features, describe them and use them to assess the validity of the algorithm response. In the work we demonstrate how the algorithm works on ID documents. Possible features are images of the coats of arms and flags of countries, background filling and text unique to the considered document type. To illustrate the proposed algorithm, the MIDV-500 and MIDV-LAIT datasets were taken. The first is used to show that the rejector does not reject correct system responses, the second – that it rejects the incorrect ones. We test several methods of zone description. The experimental results show that false type selection decreases with the use of any description type and the local CNN-descriptor shows the best performance. The increase of classes with marked zones is shown to improve the filtration of false responses. The experiments show the improvement from by 13% with one type with zones to by 4 times with 10 types.