machine-learningimage-processingdeep-learningcomputer-visionocr

Text Image Binary Classifier without deep learning or tesseract

I have 20k small label images and each image has the word "Back" or "Front".

Image resolution is all (200px, 25px)

enter image description here

I can classify these images 100% accuracy with tesseract_OCR.

        txt = pytesseract.image_to_string(img, lang='eng')
        if "Front" in txt:
            return "Front"
        if "Back" in txt:
            return "Back"

problem is, it is too slow(1 hour for 20k images) and need to install OCR packages.

I know even 3 layer Simple CNN also works well for it, but I think this problem seems to be solvable with simple algorithm without difficult techniques.

Can you recommend a new approach for me?

thank you.

Solution

  • I used the template_match function of OpenCV to compare whether the "Front" image saved as a sample matches the "Back" image.

    I succeeded in reducing the time to about 1 minute.

    https://docs.opencv.org/4.x/d4/dc6/tutorial_py_template_matching.html