How to modify the function to input multi-channel images for CNN training in MATLAB 2017a?

Hi,

I'm trying to train a Convolutional Neural Network with the newly released functions in MATLAB 2017a.

The input consists of two gray-scale images with resolution 100*100, so I stack them as a two-channel image, thus the dimension of the image is 100*100*2.

Next, I designed the architecture of input layer as

layers = [imageInputLayer([100 100 2])
          ...(omit the rest)]

However, only 1 (gray-scale image) and 3 (color image) are valid for the number of channels (the third parameter in [100 100 2]).

So my settings are invalid and the error is

Error using nnet.cnn.layer.ImageInputLayer>iCheckInputSize (line 238)
Expected input image size to be a 2 or 3 element vector. For a 3-element vector, the third element
must be 3 or 1.

I will appreciate that if anybody can help to modify the built-in function, so users can input a multi-channel image, such as a 2-channel image or a 200-channel image. I've tried but got stuck.

[morteza ali ahmadi]

I also encountered this problem. Could you find a solution?

[Zhiyi TANG]

Finally, 2018a provides the feature of multi-layer input

Answer 1

[Zhiyi TANG]

Check the doc of imageInputLayer (2018a) below:

https://www.mathworks.com/help/nnet/ref/nnet.cnn.layer.imageinputlayer.html?s_tid=srchtitle

[hiba]

please con you explain more the sysnatxe of using imageInputLayer with 3 dimention image size [240* 240*155 ] thank yoiu

Answer 2

[Muhammad Nabi]

I am having the same problem I have 4 2D feature images for the same label instance so my data set is [NxMx4] and I cant use a the CNN for it

[Zhiyi TANG]

Finally, 2018a provides the feature of multi-layer input

[Salma Ali]

Hi, I would like to use multi-input images with different sizes to CNN, is there any function in Matlab to do that? I appreciate any help. Thanks.

Answer 3

[Xin Zhang]

Hi, i am trying to iunput 4 channels (RGB+Nir) to the RCNN. Thanks to your work, it seems that the input of 4 channels is possible using the 'imageInputLayer'. But I got another question of how to make the input argument'trainingData' of the function 'trainFasterRCNNObjectDetector'? It requires as 'Labeled ground truth images, specified as a table with two or more columns. The first column must contain paths and file names to grayscale or truecolor (RGB) images'. How can we make the argument 'trainingData' as 2 channels or 4 channels? Hope to get your reply soon. Best regards,

[Buzz_root]

I have the same problem, have you found a solution?