Increasing the accuracy of evaluation of the inverse density function for a wide class of unknown distributions of measurement errors