Can you give me some advice?:)

[zoe218]

Hello, Belval, it's me again~ My mentor want me to implement crnn with tensorflow...But I'm poor at coding，I don't know how to start this project, can u give me some advice? Thanks very very very much!

[Belval]

Sure why not, as you can see this project fell way behind in my priorities but that doesn't mean I can't try to help you.

What do you need a CRNN for? Is it for text recognition or something else?

Do you have any papers on the model you are trying to implement? Aside from the one in the README.md.

[zoe218]

Thanks for replying~I want to use CRNN for laboratory sheet recognition. Now, I just want to implement crnn with tensorflow to check out if the model is suit for laboratory sheet recognition(I'm from China)or not..hah, can you understand my English?

[Belval]

Your English is very good, don't worry about it.

As for using a CRNN in laboratory sheet recognition, it probably would work but keep in mind that you'd have to extract the text zones first. You might want to look at OpenCV.

As for the actual CRNN, my project would probably work if (and that's a big if) the CTC layer worked. As of now I've had some very real trouble getting it to work and TensorFlow's documentation isn't quite reliable when it comes to recently implemented features.

Depending on what your mentor wants you might want to explore one of the following avenue:

• Figure out how to use the CTC module and fork my project.
• Use the pytorch implementation that is known to work crnn.pytorch
• Don't use deep learning as OCR can be done without it. In that case you could look at Tesseract which is a battle-tested free OCR.

While I wish I could just give you a precise estimate of when I'll be able to truly complete this. I have no way of telling.

[BeSlower]

@zoe218 You can download any workable code. Do a test on your image with their provided model first. Usually, the results are not very satisfying. Then you may train a new model with your own data. Does it have to be a TF version code? If not, you may do some experiments using pytorch version (I tested it before, and it works) as Belval said.

[zoe218]

@BeSlower Thanks a lot, :) I'll try try try!

[zoe218]

@Belval I don't know why I‘m so moved..哈哈~I'll do some trying following your advice.

[Bassmasbaa]

hello I am working on a project "text recognition with tensorflow using CNN models ! how shall i start ?? i need your advices :

[Belval]

You can simply do a letter classifier then use a feature selection algorithm + a segmentation algorithm to find the letters in your image. Then you simply stick everything together.

For example: 1- You take an image 2- You apply MSER or FAST or another feature selection algorithm (see OpenCV MSER) to get the text regions. 3- Find something to split into individual letters. 4- Run the CNN on every identified letter.

You CNN would be similar to TensorFlow MNIST tutorial one but you add letters and more training samples.

For further questions like that please contact me directly (github@belval.org) or post it on a forum, issues shouldn't be used this way.

[ajiaxiaoyi]

@Belval Hello, Belval.Can you tell me how to set this parameter 'NUM_CLASSES' in training Chinese?thank you.

[ajiaxiaoyi]

I get it.It is the number of training data types...

[Belval]

It is! Unfortunately the project isn't working right now so don't waste your time, I am still figuring out ctc_loss.

[ajiaxiaoyi]

Thanks a lot

[morrisxyang]

@zoe218 Laboratory sheet recognition have mature methods in OCR，it has reached the level of business using.After all, it's "noise" is small, and the background is not very complex.In general，CNN（or other methods in deep learning）is using for recognition in natural scene(STR)，such as the photo you taked up.I don‘t think it can help to imporve the effect and exceed mature methods in OCR. Maybe it's your boss who wants to “play” with deep learning but he actually studied other fields. Finally,I like you English and you tone,it makes me feel happy when I was debugging this project.

[zoe218]

@atVoidYX hah, thanks a lot! Your taste is great~hahah. : )

[Belval]

Closing for inactivity, feel free to reopen if you have more questions.