What counts most when it comes to recognition rate? Is it the image source or the LPR system that has the most influence over this all-important indicator – a topic of heated discussions within the LPR industry? I intended this post to be a theoretical discussion based on experience from a significant amount of cases. Therefore, I will not be the one to declare a victor but will be satisfied if I can give some new perspective on the topic.
Before moving any further, it is important to make a distinction between the following two expressions as far as license plate recognition is concerned: when I write about recognized I mean that the license plate is found in the image and the result is correct, while in case of
not recognized the license plate in the image is not found, OR a license plate is found, but the result is not correct. Both false positive and false negative results – as far as locating license plates is concerned – fall into the second category. Differentiating between these cases is out of the scope of my current post.
And now back to the question at hand! In most cases, we aim to reach the highest possible read rate of the license plates. If a license plate is not recognized correctly, the reason may be either that the input images were of bad quality or the recognition library could not recognize the readable license plate.
In the former case, we can distinguish between two disjunct sets of all events (or images): readable images (good quality events) and unreadable images (bad quality events).
This classification of events is prone to the human factor, so there can be different interpretations even for the same set of images. For now, we take these two sets as the basis of this argumentation. The number of pictures in any given set can be influenced by quality of the image source, however is definitely not the sole factor affecting it.
Let’s imagine that we process all events of both sets with an optimized LPR system. We will find that events in the “good” set AND even some events/images in the “bad” set are recognized correctly.
Now let’s take an imaginary poor performance LPR system and process the same set of images. In this case we find that some events/images even in the “good” set are not recognized. We will use the following symbols to evaluate the sets regarding how they contribute to reaching the highest possible recognition rate: neutral, good and bad.
The imaginary image processing of the two sets of events nets us 4 subsets:
correctly recognized good events
not recognized good events
correctly recognized bad events
not recognized bad events
Let’s see the findings; the subset of correctly recognized “bad”or unreadable events/images offer a nice playfield since they define a kind of “no man’s land” as far as the effect of the image quality of cameras are concerned on the process and thus characterizes the reliability of the LPR system.
Should we find too many items in the subset of unrecognized good/readable events on the opposite side of the picture: the recognition rate will be lower than expected. In this case the subset of recognized bad images will most likely contain a very small number of items.
In an ideal project, the number of “not recognized good images” is zero, and the number of correctly recognized good events is 100% of all the “good” events. Moreover, the number of correctly recognized “bad” events is bigger than zero.
What comes first into mind here is that to achieve a satisfactory recognition rate, multiple images per vehicle are required. Therefore, an algorithm is necessary for choosing and analyzing the relevant images of each vehicle that also combines their raw results into a final one.
In order to reach the final result there are 5 steps to take:
1. Choose the most relevant images from the image stream for every vehicle
2. Analyze the minimum necessary number of images to save resources
3. Analyze enough images to provide a validated result
4. Construct one result of each vehicle
5. Choose the best image of the vehicle
Considering an engine or application that uses only one image to return a license plate, the LPR system has a very important role in achieving good results even if the cameras provide high quality images. It is also imperative to understand the costs involved in improving the recognition rate. One must carefully consider if it is worth investing in expensive cameras in order to achieve the desired recognition rate or opt for an LPR system which may also produce the same results using cheaper hardware.. Asura’s ARU uses advanced video analytics that work based on an image stream instead of a sole picture of an event. ARU chooses and analyzes multiple images of each vehicle making sure that no event is missed and 98%+ is recognized correctly.
Click to see how advanced video analytics will produce results even
under difficult conditions.
HEAD OF SUPPORT
“Emergencies have always been necessary to progress. It was darkness which produced the lamp. It was fog that produced the compass. It was hunger that drove us to exploration. And it took a depression to teach us the real value of a job.”
— Victor Hugo