If you want to develop your own Convolutional Neural Network architecture from scratch and train it using MATLAB, you can use the Deep Learning Toolbox.
You can define the architecture of your CNN using the “layerGraph” object, which allows you to add different types of layers to your network (such as convolutional layers, pooling layers, and fully connected layers).
Once you have defined the architecture of your CNN, you can train it using the “trainNetwork” function. This function takes as input your training data (images and corresponding labels), your CNN architecture, and training options (such as the optimization algorithm and the number of epochs).
To answer your question about whether you have to develop the neural network algorithm yourself, the answer is no.
MATLAB's Deep Learning Toolbox also provides a variety of pre-defined functions for segmentation tasks, such as the “segnetLayers” and “unetLayers” functions. Therefore, it is not necessary to write from scratch and instead you can utilize these functions.
In terms of initializing the filters (kernels) in your CNN, you can either initialize them randomly or use pre-trained filters from a pre-existing network (such as AlexNet or VGG16) and fine-tune them for your specific task.
I hope this helps!
