2

I wanted to ask you a question about image classification. Actually I am making a image classifier and I am using convolutuional neural networks with keras and tensorflow as backend. my question is how to identify multiple objects in an image. I've tried convolutional neural network with activation function sigmoid and loss - binary_crossentropy but I am not satisfied with.

Dexter
  • 630
  • 1
  • 11
  • 24

1 Answers1

2

I'm assuming you are using a simple CNN, in which case, for multiple objects, it will only identify a single object. The problem isn't the activation or the loss, it's the architecture. You need to use something like a RCNN (Faster RCNN, YOLO, SSD, or the current SOTA: Mask RCNN).

In case you are using an RCNN model, and the detection isn't accurate enough, you would need to post more details about your network for anyone to trouble suggest improvements.

Hasnain Raza
  • 681
  • 5
  • 10
  • 1
    well I found the solution yesterday that I have to use RCNN with region of interest ( ROI) approach but I dont know the implementation so can please give me any implementation( I mean the code) of this approach – Dexter Dec 14 '17 at 15:40
  • 2
    Implementations are pretty easy to find, you just have to google them, but using them correctly requires that you have knowledge of how they work, and the libraries they are written in, one that I found easily on google: https://github.com/endernewton/tf-faster-rcnn To understand how an RCNN works, I suggest you to go through the stanford cs231n lectures on youtube, they are an excellent resource. For Tensorflow knowledge, just read the official documentation. – Hasnain Raza Dec 14 '17 at 18:01
  • https://stackoverflow.com/questions/47822154/layer-conv2d-3-was-called-with-an-input-that-isnt-a-symbolic-tensor any suggestions? – Dexter Dec 15 '17 at 04:56
  • a note for @hasnain Raza well I want to implement this in keras framework and I didnt find any link can you please suggest me something? – Dexter Dec 15 '17 at 06:13