How to write an object recognition application?

Question

Pattern recognition

I want to make an Android application in java: recognize a chair or a table in a photo.
What technologies are available for this implementation?
Read about tensorflow. In general, there is already a tensorflow light application, but I want to write it myself from scratch.
How can the model be trained? How does this process take place? Need to take 100,000 photos for sampling? Teaching a model to write in python?

Answer 1

Well, if it's from scratch, then it's easy.
1. First, let's understand what neural networks are. To do this, it’s good to study and understand machine learning algorithms in advance, and they require knowledge of linear algebra, statistics, optimization (including numerical).
On the way, we understand what is "training the model" and what is "application".
2. After that, we learn how to get an image and work with it - how to digitize an image, how to dissect it for further work.
3. After that, we begin to study the work of neural networks tailored to work with visual images - well, here you can limit yourself to convolutional networks. True, you first need to deal with other types, but I hope you can handle this easily.
4. Do not forget to simultaneously master programming tools at a sufficient level - but if you are going to write "from scratch" in java, then you will have to understand java deeply, especially since there are not so many role models in this language in your chosen area . In English - I met, but I hope this is not an obstacle for you.
To train a model written in one language, and then reprogram it to another is another challenge. But it's funny, it adds complexity to the work, increases the overall level of professional skills and abilities. So whether or not to teach Ruthon additionally is up to you.
5. Well, after that - everything is simple. We collect all of the above "up to the coupe" - and voila! - the neural network for object recognition is ready. True, after that you will have to train it - well, for this you have already prepared a couple (tens) of thousands of marked-up images of various objects, so you will not have any difficulties here.
That's about something like this. However, maybe I missed some steps along the way - so the list can expand a little. It is all the more interesting, this is why neural networks are written alone "from scratch" to develop themselves as a specialist (though it is important here that the groundwork that needs to be developed already exists, and preferably not elementary).
Better yet, find some book, fortunately now there are already plenty of them, where at least briefly the steps listed above are described - and using it as a Roadmap to move towards the intended goal, delving into the "branches" as you progress and need. At least you can hope that the author will guide you through all the necessary milestones along the way. And come to the forum with specific misunderstandings, questions. And not with "explain how this process happens, I want to write it all from scratch, but I just don't know what technologies can be used."