I'm new to object detectors and segmentation. I want to localize digits on a plate as fast as possible. All images of the dataset are normalized to $300 \times 60$. There are different approaches to solve the problem. For example, binarization + connected component labeling, vertical and horizontal projection. The aforementioned approaches fail in ambient lights, noises, and shadows. Also, there are other approaches such as STN-OCR (based on convolutional recurrent neural networks) that need a lot of plates with different composition of numbers. I have limited plates with the same numbers (about 1000 different numbers) but totally 10000 plates in different illuminations and noises. I have a good OCR (without segmentation), so I need a network just localize digits.

Is there any deep learning-based architecture for this purpose? Can I use faster RCNN? Yolo? SSD?

I trained Faster RCNN in Matlab, but it detects too many random bounding boxes for each plate. What could be the problem?

  • 1
    $\begingroup$ if you need bounding box ssd/yolo could work, if you need pixel wise precition then unet. $\endgroup$
    – juvian
    Commented Oct 21, 2019 at 16:41
  • $\begingroup$ @juvian thank you. Is there any network or architecture with less parameters? I'm worry about processing time. $\endgroup$
    – Babak.Abad
    Commented Oct 21, 2019 at 16:49
  • $\begingroup$ no hope for FasterRCNN? I'm training Faster RCNN but it seems it can not be trained very good $\endgroup$
    – Babak.Abad
    Commented Oct 21, 2019 at 17:04
  • $\begingroup$ Sorry have never used FasterRCNN so not sure about that one.I am doing japanese text detection with unet and its working fairly well. There are many posts about doing plate detection with opencv though (with some variations), maybe show some examples where it fails? $\endgroup$
    – juvian
    Commented Oct 21, 2019 at 23:20


You must log in to answer this question.