search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.79k stars 7.97k forks source link

Using yolo for license plate ocr case #4935

Open yamiefun opened 4 years ago

yamiefun commented 4 years ago

I'm working on a project that need to detect and recognize small license plates in images. The size of image is 1920 1080, and license plates in those images are about 50 30, with 7~8 characters on them.

I use yolov3 to detect license plates in those images, and works quite well. Now I can crop almost all license plates on images, so I move on to character recognization problem.

I'd like to regard characters recognization problem as a object detection problem because the font and layout of license plates are all the same, so i labeled every character on plates with seperate bounding boxes and classes(character), but this is not working. I think it's because the input images are too small (about 50*30), let alone the characters on plates.

Should I modify the config file for the neural network so it can work well on small input images and small objects? I also wonder if it's a good solution to modify 'anchors' in cfg file because the locations of characters are almost the same in every license plate image?

Thanks for all your suggestion!

stephanecharette commented 4 years ago

Like this: https://www.ccoderun.ca/programming/ml/iranian_plates.html ?

That project I did with the normal YOLOv3-tiny.cfg. No special changes needed to the .cfg file. Just lots of images.

yamiefun commented 4 years ago

@stephanecharette The size of license plates in your iamges are much larger than mine, I'm not sure if this difference caused my training failed?