Detecting objects in images using neural network

Hi,

so I'm trying to segment images to detect objects (the blobs in the attached image) by using a neural network. So I just startet with machine and deep learning and wanted to double check if my thoughts about getting started are good respectivly the right approach. So first I need to set some training examples for my network so that it knows what to search for. In my case this would be the individual blobs in my images. So I guess I have to manually extract the blobs with photoshop (something like some houndred different blobs) and train my network with it(maybe using the alexnet with transfer learning for this one class). And afterwards I should be able to run my network on a real image (my test set) like the attached one and it should detect at least some of the blobs in the image. I bet there a several steps between the input and output I didn't cover, but I just wanted to check if my approach is reasonable. Maybe there a other, better methods by using the computer vision toolbox which I don't know about. I tried using the image processing toolbox with mixed results between 50-80% accuracy, which wasn't enough and consistent for multiple different images.

So I hope someone can help me out a bit and give me some good tips and tricks. Thanks in adance!

[Image Analyst]

I don't even know what in the image you consider to be blobs. The white circles? The black debris/stains? The whole square thing? I have no idea.

[Daniel Binder]

Yeah sorry I forget to clarify that. I want to detect the white/gray circles

Answer 1

[Image Analyst]

You should have a template for the aligned image that specifies where each circle is. Then rotate your image so that it is aligned, like with imregister(). Then you can simply use your template.