Is it possible to write a CNN for detection of the objects?

i.e. without the introduction of ready trained models. For example, you need a photo to recognize several types of wastes: metal, paper, etc. Or to use a ready-made architecture and to teach? how difficult would it be to implement it, if you do in python with keras? I need to recognize only the types of wastes that nothing else but them.
I would be happy if I showed the architecture of such a network

p.s. question thesis
April 7th 20 at 15:34
2 answers
April 7th 20 at 15:36
It is possible and from scratch. The question is, do you have enough knowledge, abilities, skills, experience, time and other resources. Judging by the fact that you asked this question, you will be able to do. Even more so, in the framework of the diploma.
As for the "show architecture"... of course you can, but good.
Well, for example:
April 7th 20 at 15:38
You can take the ready model, already trained on a large set of images (e.g. for ImageNet)
and retrained to classify your images.

Here's the article (in English). with reference to laptop Colab'e, where we take EfficientNet and learn to distinguish cats from dogs.

I've read that there are R-CNN, faster R-CNN. If I instead of adding ready-made models (for example VGG16) in the new architecture model.add(VGG16()), just add its entire architecture thus
etc. and import weight whether work correctly or something extra need to do?
and the second question: to teach the network to detect multiple objects in one photo, enough to dataset consisted of images of one object or needs to be multiple objects on one picture? - ruthie_Denes commented on April 7th 20 at 15:41

Find more questions by tags Machine learningNeural networksPython