I would like to suggest you first to understand it conceptually : https://medium.com/analytics-vidhya/convolutional-neural-networks-cnn-explained-step-by-step-69137a54e5e7
Hi, if you cannot find exactly what you thought you will get, then its a good research question to answer.
a bit of guideline that can help based on my understanding
1. Group the applications or problem that computer vision scientist always aims to solves such Segmentation problems, classification and image-to-image translation. You can then find a survey or review papers based on the particular problem and see how researchers are weighting the solution of deep learning(CNN) vs traditional techniques which will then give you better pros and cons with evidence.
2. You could then select review and survey works from different application domains where these techniques(deep learning ) are been used such as in face recognition, biodiversity applications, human activity recognition and many other domains. This will give you a picture of how researchers of that particular domain of application are adapting the techniques.
Attached is one of the article that you can start with
1. Article A Survey on Deep Learning: Algorithms, Techniques, and Applications
As far as I know convolutional neural networks are usually not trained while processing a hitherto unknown image. Consider a medical image with a structure in it. In order to do a segmentation, an algorithm might start on a border and follow it. On its way, there is a left area and a right area. Their properties like texture, and color, may change gradually in an unforseeable way. An algorithm can cope with that by updating its internal state, but a convolutional neural network is not designed to do this.
Or imagine an image with polygons on it. Even if some lines might be blurred and portions of them missing, an algorithm could be able the reconstruct them by knowing about the nature of the underlying polygon, as opposed to convolutional neural networks.
Attached is an image with an enclosed area. The problem is now to classify the blue, red, white, green, and yellow square as being inside or outside. With an algorithm this would be no problem: it would count the number of border crossings starting at a square and marching away. But a convolutional neural network has no built-in capability of executing simple algorithms like counting events.
Not sure this directly relates to your question, but here is a database of some papers on CNNs every data scientist might find helpful. https://towardsdatascience.com/5-papers-on-cnns-every-data-scientist-should-read-c689b686aa83
in this (open access) paper: Article Enhanced robustness of convolutional networks with a push–pu...
we discussed the problems of robustness of CNNs to corruptions/perturbations of the test images that have not been seen during training, and proposed a new type of layer that improves the generalization.