Making and training a Neural Networks model (using keras) to recognize handwritten digits.
Currently the model gives an accuracy of 98.18% which can be improved upto 99.8% using more complex Convolutional Neural Networks(CNN). Right now I've used a really simple neural network since I've got to learn more about CNNs. So, I'll be improving this project in the near future.
- First image in the dataset
- Second image in the dataset
pandaskerasnumpymatplotlib
When you run the program using python trainer.py you should see the following if you already have the above libraries installed.
- Output
You'll still have to feed a 28 by 28 pixel image to this model since it's not a robust model and does very little pre-processing on it's own. The images in the dataset are in the form of a 2D array which is stored in numpy arrays. I hope to improve the model itself as well as the pre-processing in the future.
I have now deployed the model on web using Python's backend framework Flask
opencv(imported ascv2)tensorflowflask
-
The api.py makes the server and runs as the backend program. This takes in the image file from the form and uses them for prediction. To run the api.py head to the command prompt, change the directory and type python api.py.
-
Open the web-browser and go to http://localhost:8001/ to get your index.html running.
For now I have uploaded two pictures from the dataset itself that you can use for testing the deployed model. pic1.jpg has the number 5 and the pic2.jpg has the number 0.
- index.html
- predict.html
In the future I'm hoping to add more pre-processing for predicting the pictures that are not present in the dataset. Will most probably use thresholding on the images based on colors using OpenCV library.