How to create the groundTruth and evaluate the classical computer vision model for crack detection and does training models only restricted for machine and deep

If I have a model to detect cracks which can be consider as classical computer vision model, it works based on thresholding and morphological operation to detect the cracks on the surface, the model is based on Matlab, My questions are : 1- How to create the ground Truth and it necessary for this type of model for performance evaluation? 2- How to evaluate this model based on precision and recall if I have 100 images or more? 3- Do I need to train this model or training restricted to machine learning and deep learning models only? I attached the original image (The model input) and the model output