CNN-Model Training for Symbol Detection.

[Madacool01]

The next major step for my Air-Math project will be training a CNN (Convolutional Neural Network), whose job is too look at the pixels from my webcam, draw boxes around them in its mind, and guess what they are. Here's how it works:
1. Scanning the Paper:
The CNN starts by sliding small "filters" (tiny 3x3 pixel squares) across my image. This process is called Convolution. These look for specific patterns. As the image moves deeper into the network, the layers work like a team:

• First Layers: They look for basic lines. They notice where the pen strokes start, stop, or curve for example.
• Middle Layers: They put the lines together. They recognize shapes such as "2 crossing lines $+$", or "a closed circle $0$".
• Deep Layers: They look at the big picture. They realise that a circle on top of another is an "$8$" for example.

2.Finding & Naming:
Every single "filter" square on the image asks itself 2 questions:

• "Is there a math symbol centered inside me?" If yes, it makes a digital box around the drawing.
• "What symbol is it?" It looks at the shapes inside its box and calculates the percentage $P(A)$ (e.g 95% sure its a 5, "3%" sure its a S, "2%" sure its a 2 )

I will be training this model on a Dataset from Kaggle inside of Google Colab because of its free NVIDIA T4 GPU