Browse Prior Art Database

Efficient License Plate Recognition Using Image Quality Information

IP.com Disclosure Number: IPCOM000237922D
Publication Date: 2014-Jul-21
Document File: 7 page(s) / 150K

Publishing Venue

The IP.com Prior Art Database

Abstract

Automatic license plate recognition (ALPR) systems are evaluated using various criteria, three of which are yield, accuracy, and execution time per image. Higher yield and accuracy are desired in order to minimize the number of images that are sent for human review but this often comes at the cost of higher execution time since multiple and/or more complicated and time consuming algorithms must be deployed to ensure an accurate, automated conclusion. In order to keep up with traffic volume, higher execution time leads to the requirement of more processing power in the form of additional servers, cpu-cores, and power consumption. This problem is most acute for poor quality images where the ALPR engine struggles to reach a good conclusion and in turn spends disproportionately more time before returning a non-conclusion. Given the problem of high execution time for poor quality images, we propose a method of assessing image quality early in the ALPR process and terminating execution when certain criteria is met indicating low probability of a successful conclusion. We've identified four image quality metrics that we use as features to train a classifier to determine whether the ALPR engine is likely to conclude on an image given its image quality. The classifier is effective at identifying no-conclude images but not perfect and we provide a parameter to control the balance of lost yield vs. execution time savings given business need.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 28% of the total text.

Efficient License Plate Recognition Using Image Quality Information

Automatic license plate recognition (ALPR) systems are evaluated using various criteria, three of which are yield, accuracy, and execution time per image. Higher yield and accuracy are desired in order to minimize the number of images that are sent for human review but this often comes at the cost of higher execution time since multiple and/or more complicated and time consuming algorithms must be deployed to ensure an accurate, automated conclusion. In order to keep up with traffic volume, higher execution time leads to the requirement of more processing power in the form of additional servers, cpu-cores, and power consumption. This problem is most acute for poor quality images where the ALPR engine struggles to reach a good conclusion and in turn spends disproportionately more time before returning a non-conclusion.  Given the problem of high execution time for poor quality images, we propose a method of assessing image quality early in the ALPR process and terminating execution when certain criteria is met indicating low probability of a successful conclusion. We've identified four image quality metrics that we use as features to train a classifier to determine whether the ALPR engine is likely to conclude on an image given its image quality. The classifier is effective at identifying no-conclude images but not perfect and we provide a parameter to control the balance of lost yield vs. execution time savings given business need.

Background:

ALPR systems are structured as shown in Figure 1. The first step given an input image is to locate candidate plate regions of interest (ROIs). For a tolling application, the input image typically contains no more than one license plate but for other applications or for cases where multiple lanes are imaged by a single camera, multiple plates may exist in the input image. For plate localization, we typically bias on extracting many ROIs to ensure that at least one ROI includes the license plate image.

The set of ROIs is passed to 'Character Segmentation' where the characters are normalized and segmented before being passed to OCR. As part of the normalization and segmentation process, many algorithms are employed to ensure good segmentation such that all characters are cropped effectively; non-characters and border artifacts are excluded. OCR generates the most likely code conclusion given the segmented character images and the code along with OCR font is used to determine the most likely state.

When profiling ALPR performance over 1000's of images, we've learned that the segmentation subsystem typically consumes the most time. Depending on the image set and parameter settings, this can be as high as 80% of the total image processing time. This is driven by several factors one being the number of regions returned by localization and the second being various rules, one of which is that license plates typically have at least four characte...