Is there any way to use deep learning model for higher resolutions images (like 1024x1024)

I want to use mobilenetv2 model with transfer learning. It uses 224x224 image resolution in default for the image classification However I use a high-quality camera for image detection and it gives High-quality images like 1024x. I can use the imresize function to convert the 1024x to 224x. But It seem like it lose the information of the input image and I tried to change the input layer of the model to 1024x. But it gives an error when I replace with this layers.Is there any method for this,Thank you

Answer 1

[Tomaso Cetto]

Hi Hasintha,

When confronted with this issue, the recommended workflow is usually just to resize the images. That said, the workfow you suggest might work depending on what your end goal is.

You should be able to remove the input layer of MobileNetv2, drag a new imageInputLayer, define the correct input size, and re-connect that new input layer to the first convolution layer. If you have any problems doing this, do let me know exactly what the issue/error is, and we can take it from there!

Once you have the new network in DND, you should be able to export it back to the workspace, and re-train on your new images. However, because the network has been trained to process images which are much smaller than your new ones, there is no guarantee it will behave exactly as expected. Again, if this doesn't work, please let me know what error you see and I can help you fix it!