Neural network-based prediction of the stopping moment for text recognition in a video stream

This paper considers the problem of finding the stopping point for the process of text recognition in a video stream. An overview of existing approaches is presented, which are based on modelling and thresholding of the expected recognition error. As an alternative, this paper proposes a neural network-based method. The method uses a fully-connected network that takes a sequence of normalized Levenshtein distances between the consecutive combined recognition results of several previous frames and predicts the pseudo-probability of stopping on the current frame. A process of data preparation and training protocols, as well as the parameters of the model, are presented.Experimental evaluation of the proposed method shows that the learning-based approach allows to reduce the mean recognition error when compared to the base stopping method in every tier of the mean number of frames, with a maximal normalized Levenshtein distance improvement of 12.7% with not more than 10 processed frames on average on MIDV-2020 dataset and 15.4% with not more than 17 processed frames on average on the test subset of RoadText-1K dataset.