How can i customize my Convolution Neural Network (CNN) to deal with gray images (2D ultrasound) as the input layer is something like (3, 256, 256) the 3 represents the R, G and B channel in CNN but in my case the number of channels will be only one.
when i run the code i got the error :
ValueError: Cannot feed value of shape (16, 256, 256, 3) for Tensor 'x:0', which has shape '(?, 256, 256, 1)'.
is there any workaround for this without affecting the network performance?
Amr,
intensity images (grayscale) correspond to color images with equal values in all three color channels. Export your grayvalue images to color RGB images with any conversion software such as Gimp or Linux command line packages such as ImageMagick (convert command). If you're a Windows or MacOS user, you can use Adobe Photoshop (Image -> Mode -> RGB Color) or any similar thing. You'll obtain an image where all channels have equal values and the CNN will be able to load it.
Another, much different question, but that I think is more important in your case, is if the CNN you're using is adequate for this task you intend to perform. Which CNN are you using?
Thanks Aldo for the reply.I will try to use what you are suggested.
Regarding The CNN i use, its built using tensorflow, i want to extract the last layer features.
I would recommend the quick version:
https://stackoverflow.com/questions/40119743/convert-a-grayscale-image-to-a-3-channel-image
regrads
What I dont understand about the solution presented here is you are asking that 3 channels which presumably would be identical are submitted to the model instead of using a model that accepts the input of one channel and as such would be less computational. Sorry but still learning this stuff
Here is the complete solution : https://towardsdatascience.com/autoencoder-for-converting-an-rbg-image-to-a-gray-scale-image-3c19a11031c9
Amro Saleh there is no need to convert a gray scale image to RGB, it will definitely increase your computation cost as Ken Mascola says. You can directly pass a grayscale image to CNN by setting your channel dimension as 1. Check this link:
https://yashk2810.github.io/Applying-Convolutional-Neural-Network-on-the-MNIST-dataset/
Your problem is as simple as working in a MNIST dataset that has only grayscale images.
It depends on type of dataset you use since you dataset has ultrasound image so their is uniformity among data channel and features will be more patterns and shared but color will not add any to you model, now I understand the model you are referring to is using 3 channel and simple fix to go to staring input layer and replace 3 by 1 and similarly also check wieghts matrix should be initialise accordingly. In short CNN can be trained on 1 channel as well generally people use that as ANN or FCNN on MNIST Dataset but you can run convolution on 1 channel create features, flatten it and dense layer
- Badges
- Science topic