validation function and input size change for mask r cnn

Hello, i have a question about mask r cnn training process.

thanksfully, i could try to use the mask r cnn to detect particles in this images because this site explained about mask r cnn in matlab.

https://kr.mathworks.com/help/vision/ug/example-InstanceSegmentationUsingMaskRCNNDeepLearningExample.html?searchHighlight=mask%20r%20cnn&s_tid=srchtitle_mask%20r%20cnn_1

However, during the result of training does not works well and i realized that the site does not mentioned about validation during the process. is there any way to use validation?

Also, as you can see, my images are gray scale. Can i used the image as n*m*1 not n*m*3 as RGB? it looks RGB needs much high computation power.

Thanks.

Answer 1

[Avadhoot]

Hi Seon Je Kim,

I have inferred from your first question that you want to use validation in the training process of the mask RCNN. There is a way to do that. You could just specify the parameter called as ValidationData in the trainingOptions function for your neural network. The validation data could be a table, a datastore or even a cell array.

To obtain the neural network that performed best on the validation dataset, you can set the OutputNetwork parameter to "best-validation-loss".

More information regarding validation can be found on the following resource : Validation in trainingOptions

Regarding your second question, if the images are all grayscale, you can reduce the number of channels from 3 to 1. You could specify that in the code as:

imageSizeTrain = [n m 1];

Before making this change, please ensure that you won't lose valuable information by converting the images to grayscale, as it could adversely affect the model's performance.

Hope this helps.

Regards,

Avadhoot

Answer 2

[Ashish Uthama]

Thanks for posting the image!

You could try the Cellpose add-on and explore different models to see if some of the canned ones get you close to the results you want. Else, you could use these labels as a starting point to train your own cellpose model.