recognition improvement about keras-ocr HOT 4 CLOSED

First, thanks for trying out the package!

Getting detection to work well in this case is relatively easy but improving recognition will take more work.

On detection, because the text is so small upscaling is very important. The default upscaling of 2x works but because the image is so tall, you hit the default max_size limit too soon. I found setting that parameter to 4096 to be enough to get high quality text detection.

import matplotlib.pyplot as plt
import keras_ocr

image = keras_ocr.tools.read(
    'https://user-images.githubusercontent.com/25033310/74313656-fece6500-4d6b-11ea-8fae-082421897241.jpg'
)

pipeline = keras_ocr.pipeline.Pipeline(max_size=4096)
predictions = pipeline.recognize([image])[0]

fig, ax = plt.subplots()
keras_ocr.tools.drawAnnotations(image=image, predictions=predictions, ax=ax)

The bounding boxes look pretty good -- but the text recognition is quite poor. To fix this, I think you'll need to get a dataset together of text that looks like that in your use case to train a custom recognizer. The good news is that it is much easier to train the recognizer than the detector! There's an example for fine tuning the detector in the documentation which I think you may want to adapt.

I'm going to close this now because it is not so much a bug in the package as it is a how-to question which I believe my comments above answer. If you run into problems with the example or with training the recognizer in general, please open a new issue.

from keras-ocr.

I forgot to run your second example, but that one actually does a little better.

import matplotlib.pyplot as plt
import keras_ocr

image = keras_ocr.tools.read(
    'https://user-images.githubusercontent.com/25033310/74313659-ffff9200-4d6b-11ea-907e-89bd0459f275.jpg'
)
pipeline = keras_ocr.pipeline.Pipeline(max_size=4096)
predictions = pipeline.recognize([image])[0]
fig, ax = plt.subplots()
keras_ocr.tools.drawAnnotations(image=image, predictions=predictions, ax=ax)

from keras-ocr.

from keras-ocr.

Yes, that's correct -- it appears that you will need to obtain some real-world labeled data to train the recognizer for this use case.

from keras-ocr.