What Kind of Pre-processing is required for images for transfer learning and training AlexNet?

Hello,

I am working on a CNN model for classifying pavement distresses and I am using transferred learning through alexnet. I was wondering if my images require any kind of pre-processing such as (semantic segmentation) other than resizing them. I am using high-resolution colored images of pavement distresses.

Thanks in advance.

Answer 1

[Raunak Gupta]

Hi,

There are several preprocessing techniques which can be used for preparing data to pass to a network for transfer learning applications. Though I cannot suggest a specific step for your case, you can refer to the following documentations which explain more about these techniques.

https://www.mathworks.com/help/deeplearning/ug/preprocess-data-for-domain-specific-applications.html

https://www.mathworks.com/help/deeplearning/ug/preprocess-images-for-deep-learning.html

[Vinay Chawla]

Thank you.

[Vinay Chawla]

I get that I will be required to do the resizing to fit the network requirements. But, do I have to pre-process images (e.g. contrast enhancement, sharpening. noise removal,etc) before training?

And Is it necessary for me to perform these pre-processings before the training?

[Raunak Gupta]

Hi,

The pre-processing is usually done in cases where the training results on unprocessed data are not good. In those scenario pre-processing can make images more suitable for training. If you are getting good results without using any pre-processing techniques its not necessary to apply any of them.

[Vinay Chawla]

Thank you so much.