Limitations of image size for differnet networks, for image classification? I want to differenciate between screws and rings and the image size that I have is 400*400. I wanted to know which network would be suited for such size and or there is a limitation like for every network for example for MLP you cant have a Image size greater than 128 *128. For LEnet you cant go above 300*300?
Please post also the helpful resources.
Thank you
Just change the number of layers and size of filters of CNN and monitor the results until you get perfect accuracy
Pretty much what Faheem Khan said. Architectures are not limited by resolution per se. You can, for example, modify pooling layer to increase downsampling factor, or adding more layers.
Example: 400x400 input -> conv+maxpool (2x2 window, 2x2 stride) -> 200x200 -> conv+maxpool (2x2 window, 2x2 stride) -> 100x100 -> conv+maxpool (2x2 window, 2x2 stride) -> 50x50-> conv+maxpool (2x2 window, 2x2 stride) -> 25x25x number of feature maps -> vector with (25*25* number of feature maps) elements.
Of course pooling can be switched out for some other perhaps more suitable downsampling method, but you get my point.
Happy training :)
- Badges
- Science topic
- Similar topics
- Computer Science
- Images