unable to recognise text with numbers

[p9anand]

Hi,

Thanks for this amazing work. I was using this repo for recognition task but was unable to do that. Attached is the image. Please help me how i can do that. It detected it as: alainatism

I used this model.
CUDA_VISIBLE_DEVICES=0 python3 demo.py
--Transformation TPS --FeatureExtraction ResNet --SequenceModeling BiLSTM --Prediction Attn
--image_folder demo_image/
--saved_model TPS-ResNet-BiLSTM-Attn.pth

[ku21fan]

Hello,

In my opinion, the problem is 'noise' rather than 'number', since this example has heavy background noise.
Can you try with

1. other examples which did not contain numbers?
2. other examples which contain numbers with low background noise.

Best

[Ameya-Manas]

Hello @ku21fan

On the repo home page, MJSynth (MJ) and SynthText (ST) are mentioned as the training datasets. On the SynthText (ST) website, the following image shows that the model has been trained on such images:

This shows that the model has been trained for digits as well.

The results for some images with digits is, however, not correct. The images are listed below.

The first image gives the correct prediction. The seconds and third image however give the results "be" and "peg" respectively.
Can you please suggest as to what might be the issue or what I might be doing wrong? The demo.py is being run for the best accuracy model shared in the repo.

Best.