2

I'm using a Faster RCNN network to perform object/symbol detection and I'm facing 2 major issues.

  1. The bounding box of the detected symbols is not tight enough. For example, in many cases, only 50%-70% of the entire symbol is being identified (Example: Resistor R1 in the image below). What can I do to make my bounding box more accurate?

  2. In the below example we have 3 resistors, R1, R2, R3. The trained network is able to identify R1 with partial IoU, R2 properly but it has missed out R3 completely even though R3 is present on the same page and is the same symbol as R1 and R2. Why does this happen and how can I overcome this? (I tried a correlation-based approach but there are too many variations to consider in my use case)

How can I fix the above issues? Thanks in advance.

Example Symbol detection on test image

Abhishek Sudhaker
  • 31
  • 3
  • Did you train by yourself or you used pre-trained data set trained by someone else? If your sample images will be clear like this, why you not try to use morphological operations? – Yunus Temurlenk Jul 24 '20 at 05:18
  • Yes, I did train the network myself using custom data. This sample/representative image happens to be super clear, but generally the images/symbols have a lot of noise and we are looking at nearly 30-35 symbols to be identified on a page. – Abhishek Sudhaker Jul 28 '20 at 09:07
  • If you can clear the noise for the other samples, I am thinking morphological operations will be more helpful – Yunus Temurlenk Jul 29 '20 at 04:33

0 Answers0