Is it possible to write a CNN for object detection from scratch?

Question

Python

those. without the introduction of ready-made trained models. For example, you need to recognize several types of waste in a photo: metal, paper, etc. Or take a ready-made architecture and teach it yourself? How difficult would it be to implement this if done in python with keras? I need it to recognize only waste types, nothing else but them.
I would be glad if they showed the architecture of such a network

ps question of the thesis

Answer 1

It is possible from scratch. The question is whether you have enough knowledge, skills, experience, time, and other resources. Judging by the fact that you asked this question, you will not be able to do this. Even more so, within the framework of the diploma.
And as for "show architecture" ... of course you can, but it's useless.
Well, for example:
https://www.kdnuggets.com/2019/08/2019-guide-objec...
https://towardsdatascience.com/face-detection-with...
https://www.analyticsvidhya. com/blog/2019/07/compu...
https://towardsdatascience.com/real-time-object-de...
https://towardsdatascience.com/detailed-tutorial-b...

Answer 2

You can take a ready-made model already trained on a large set of images (on ImageNet for example)
and retrain it to classify your images.
Here is an article (in English) with a link to a laptop in Colab, where they take EfficientNet and learn to distinguish cats from dogs.

Architecture

отсюда: