Digit recognition using PyTorch and neural networks I also created a script which resize every image thay user upload into program into 28 x 28 px .png file Here are some examples of how it works:
It trains on special dataset from MNIST library and then calculates the result after each epoch
Example for number '2' from MNIST library
I tested it also in not easy examples such as Captcha numbers
Example for number '2' from random Captcha screen
For example below it didn't work because of the quality after program resized the .png image
Example for number '6' from random Captcha screen
I tried also the "hand made" examples that I did in paint and here is how it looks:
Example for number '7' from paint
Program is working with around 97% accuracy for MNIST library and 90% of the time for other .png (I tried 10 examples) When it comes to regular square .png images, the program handles them well. However, issues can arise when working with non-square .png files, such as one with a resolution of 200 x 800 pixels. In such cases, the program may produce unexpected results. While a perfect square isn't necessary, the image should have dimensions that mimic a square to ensure the program works as intended.