I am developing a project where the objective is to design a neural network that recognizes Ecuadorian license plates in real-time. I already saw some projects on this topic, and I found the following project where the author developed a neural network in TensorFlow (http://matthewearl.github.io/2016/05/06/cnn-anpr/ ). I did some adjustments, but I don't know if this is the best option. I would like to compare with other already known methods. Somebody could help me with some information about that?
P.D.: I am also open to information that allows me to design the network by myself.
Hi Oscar,
The sliding window approach shown in the link you provided is definitely the enemy of "real time" notion. I suggest examining the latest version of YOLO (You Only Look Once), it is a real-time object detector and has won many competitions. I understand that you are after recognizing car plates, not recognizing the characters on the plate, because that is another story.
Please have a look at this method. This was the SOTA method some time ago:
https://paperswithcode.com/paper/lprnet-license-plate-recognition-via-deep
Also, do have a look at Ceternet for the same. I worked on it during one of my projects and the time was good as compared to some other methods.
Or you can always go with a fast inference model at the cost of a little accuracy. Like Mobilenet, etc.
Thank you for the information Ashay Singh and Aparna Sathya Murthy. I truly appreciated all your help, I will review something about that too.
- Badges
- Science topic