Double digit recognition
Hi,
I would deeply appreciate a high-level input on how to go about digit recognition for my dataset.
Below are two images from the dataset.
The numbers shift on the image when they go from single do double-digit, so I cannot simply crop each number by hardcoding its coordinates.
Therefore I need to:
- To isolate each number
- Pass each number for recognition.
What would be the right approach for 1. and 2. for this dataset?
Thanks!
what did you try, on your own, already ?
Hi @berak,
I did try tesseract but it is not working well (can't restrain it to only digits). Also, I am thinking there must be a simpler solution to this dataset.
I am thinking of cv2.findCountours to isolate the digits so I can pass them to a classifier.
When it comes to classifiers I am familiar with NN variants (trained on handwritten digits). So one way is to take a trained NN and finetune it to this data. I am also thinking about is KNN (nearest neighbor).
Generally, I am looking for advice which classifiers would make sense for this dataset. Then I'll compare them and pick the fastest. Accuracy should be 100%. Thanks.
I find recently people in the forum like to use tesseract. Why would people tend to choose this method these days?
as far as I am aware, outside of building your own OCR classifier, there really is not that many options for opensource OCRs. you can either use tesseract (and its been upgraded with newer LSTM based algorithms) or use services such as amazon rekongiton,...
If the digits don't vary too much in scale or rotation, we find template matching to work well.