Convolutional neural networks, Convolution Core Matrix

Hello! In the very first stage, when we get the original image (for example, a photo of a letter), we need to go through it with a window of n*n size and multiply by the kernel (convolution matrix) to build Feature Maps. But it is not written anywhere what values should be in the core itself (the convolution matrix, that is, what values should be multiplied by the matrix), whether it will be possible to use this matrix as the convolution core to determine edges?

Also, if the size of the input image is 30*30, then will it be possible to pass through it with a 5*5 window, this is enough to achieve optimal recognition accuracy?

By which convolution kernel is it best to multiply a section of the input image for the highest recognition accuracy? Or are all the values in the kernel matrix initially zero? Is it still possible to ask by what rule or formula the number of feature maps is determined? Or if the task is in if you recognize 26 letters of the English alphabet, then there should be exactly 26 of them in each stage of building feature maps? Thank you in advance!