98 points by deeplearning_enthusiast 6 months ago flag hide 22 comments
johnsmith 6 months ago next
Impressive work! I've been looking for a good solution for signature verification. Going to try this out with my own data.
anonymous 6 months ago next
Have you considered using PyTorch instead of TensorFlow? In my experience it's faster and easier to work with.
johnsmith 6 months ago next
I've used PyTorch in the past and found it to be slower for this particular task. TensorFlow's optimized C++ core gave me better performance.
techgirl 6 months ago prev next
I love seeing HN posts about machine learning and computer vision. This is fantastic work!
curiouscoder 6 months ago prev next
What's the accuracy of your handwritten signature verification model?
johnsmith 6 months ago next
The model has an accuracy of 98% on my test dataset.
mlwhiz 6 months ago prev next
That's really impressive. How many signatures did you use for training and testing the model?
johnsmith 6 months ago next
I used 5000 signatures for training and 2000 signatures for testing the model.
machinelearningpro 6 months ago prev next
Awesome work! Any chance you can open source the code and the dataset? It would be really useful for the community.
johnsmith 6 months ago next
I'm planning to open source the code and a portion of the dataset in the next few weeks. Stay tuned!
johnsmith 6 months ago next
Thank you! I'll post an update when the code and the dataset are ready to be released.
somebodyelse 6 months ago prev next
Can't wait for the code and the dataset to be open sourced. This is really interesting work!
newcoder 6 months ago prev next
Great job! Will this work for approving financial transactions as well?
johnsmith 6 months ago next
I think this model can be useful for financial transactions, but further testing is required. Keep in mind that signatures can be easily forged, so I would recommend implementing multiple authentication factors for financial transactions.
alertuser 6 months ago prev next
Be careful not to use this for critical applications, as there have been multiple cases of deep learning models being fooled by adversarial attacks.
johnsmith 6 months ago next
You're right. Adversarial attacks can be a real problem in deep learning models. I'll add a note in my documentation and suggest using additional authentication factors.
curiousgeek 6 months ago prev next
How did you design the architecture of the CNN?
johnsmith 6 months ago next
I used a standard GoogLeNet architecture with some modifications like batch normalization and dropout. I also replaced the last FC layer with a sigmoid layer for binary classification.
technicaluser 6 months ago prev next
Can you share your training logs and hyperparameters?
johnsmith 6 months ago next
Sure, I used learning rate of 0.001, batch size of 512, and a total of 20 epochs. I'll post the training logs and the config file soon.
userwithquestion 6 months ago prev next
What kind of preprocessing did you do on the signatures?
johnsmith 6 months ago next
I applied some basic image preprocessing techniques like binarization, noise reduction, skeletonization, and segmentation. I'll write a separate blog post on the preprocessing methods soon.