On optimal stopping strategies for text recognition in a video stream as an application of a monotone sequential decision model

Abstract

The paper describes the problem of stopping the text field recognition process in a video stream, which is a novel problem, particularly relevant to real-time mobile document recognition systems. A decision-theoretic framework for this problem is provided, and similarities with existing stopping rule problems are explored. Following the theoretical works on monotone stopping rule problems, a strategy is proposed based on thresholding the estimation of the expected difference between consequent recognition results. The efficiency of this strategy is evaluated on an openly accessible dataset. The results show that this method outperforms the previously published methods based on identical results cluster size thresholding. Notes on future work include incorporation of recognition result confidence estimations in the proposed model and more precise evaluation of the observation cost. Keywords Recognition in video stream • Mobile OCR • Stopping rules • Decision making • Mobile document recognition • Anytime algorithms This work was partially financially supported by Russian Foundation for Basic Research, Projects 17-29-03170 and 17-29-03370.

Monitored Reconstruction: Computed Tomography as an Anytime Algorithm

“…Anytime algorithms as a way of thinking about algorithms with quantifiable goals are useful when the cost of computation (whether in terms of time or other quantities) is comparable, or at least relevant, in relation to the cost of error. Intelligent systems such as decision support systems [36] and computer vision systems [37]- [39] use the model of anytime algorithms to represent and manage the trade-off between the quality of the result and the time required to obtain it. If the tomographic procedure is not broken down into separate stages scanning and reconstruction, the tomographic scanning cost can be either expressed in terms of time required to collect the projections, or in terms of radiation dose delivered to the object.…”

“…For the case of monotone stopping problems where the error term of the loss function (1) is expressed as a distance ρ from an obtained result to the ground truth, i.e. (R n , θ) = ρ(R n , R * (θ)), in the paper [37] an approximation of the myopic rule (11) is proposed. Instead of estimating the difference between the current error and the expected error at the next stage it is proposed to estimate the expected distance between the current result to the result which would be obtained on the next stage.…”

“…Using triangle inequality it can be shown that under the hypothesis H2 at the stage when the stopping conditions for the rule (13) are satisfied, the stopping problem becomes monotone starting from this stage and it is optimal to stop [37]. The rule (13) then takes the following form:…”

“…4) For implementation of the stopping rule (16), the value of the Radon invariant S(R T ), which can be calculated by analyzing the obtained projections x 1 , x 2 , ... , x n . The method of modelling of the next result introduced in [37] is not applicable for the case of tomography (as the assumption of the next projection x n+1 having the same value as one of the previously acquired will lead to the same reconstruction result), thus in order to estimate the expected distance E n ( R n − R n+1 2 ) other methods should be used, such as methods of time series forecasting. We used the most basic estimation method where the target expected distance is assumed to be close to the distance between the two most recently obtained results:…”

Next integrated result modelling for stopping the text field recognition process in a video using a result model with per-character alternatives

“…This method, however, is not fully formalized and raises the questions of tuning the clustering parameters. In [15] [3] video stream recognition stopping as a monotone sequential decision problem. It presents a stopping strategy derived from the properties of monotone stopping problems, however it was tested only for text recognition results as simple strings, without any per-character alternatives.…”