how to combine two separately trained semantic segmentation neural networks into one to detect mixtures

I'm using machine learning to train a semantic segmentation network model to segments certain objects in image files. I have different objects in my image files that i want my network to be able to segment and differentiate. For now i have been abe to separately train networks for each of them and these networks can segment each of the objects in test images containing the specific objects. However i want t have a network that can detect each object in an image containing all 3 types of objects that i want to differentiate. How do i go about this? Can i combine the 3 separate networks i have individually trained to be able to achieve this? What will be the best way to do this. @MathWorks Support Team

Answer 1

[Mahesh Taparia]

Hi

You can train a single network which can segment the multiple classes in an image. Assign the different label ID of each class and train the network. Then evaluate that trained network on the test dataset. You can refer this documentation of multi class semantic segmentation using deep neural network. Hope it will help!

[Brian Chesney Quartey]

follow up: what is the best way to create the pixel label datasets? Is there a faster and automatic way to do this? I have a large number of datasets and cannot use the image labeler to label all of this. I tried using treshold segmentation to perform the labelling for each image dataset type. but that doesnt segment into different classes.

For the documentation you just suggested that i look at, it seems the image dataset had all classes in one image. For my dataset, different classes are in different set of image datasets. Whats the best way for me to go about this?

[Mahesh Taparia]

Hi

You can read the multiple images at a single go and store the labels into new variable and save that as label image.

For example, suppose if you have 3 label images ( I1, I2 and I3) of 3 different classes of the same scene., then read these label images and assign the different label and store it. For example:

newLabel = zeros(nRow,nCol);
newLabel(I1==1)=1;  %Label ID 1
newLabel(I2==2)=2;  %Label ID 2
newLabel(I2==3)=3;  %Label ID 3
%save this newLabel as png image and then create the datastore and follow the example.

For multiple images, take a loop, perform these operation and save label image for each file.

[Brian Chesney Quartey]

I'm not sure if i understand this

[Mahesh Taparia]

Hi

In your older comment, you mentioned that you have the labels information of one class in one label image, i.e if the image have 4 classes then you have 4 different label images.

You can combine this label information into 1 image by mapping those label value to 4 different categorical labels (say 1,2,3,4) which is also followed in this documentation. My previous comment was just a sample illustration.