A method is disclosed for reading license plate numbers in a road network, comprising: recording an image of a license plate number at a first location, OCR-reading a license plate character string in the image, and storing an OCR data set in a database; recording an image of a license plate number at a second location, OCR-reading a license plate number character string in the image, and generating a current OCR data set; and, if a confidence measure of the current OCR data set falls below a first minimum confidence value, selecting a stored OCR data set with a license plate number image having a similarity exceeding a minimum similarity value and/or having the greatest similarity to the license plate number image of the current OCR data set, and using the selected OCR data set for improving the license plate number character string of the current OCR data set.