Different filters for pretrainned network

if i have a group of images and i filtered all of them with different filters , is there a way or method or a published something or calculate a fator to know me which filtered group is best with specific pretrainned transfer learning network for classification task using matlab without training the network with each group of filtered images

[Dina]

could you help in this @Image Analyst

Answer 1

[Image Analyst]

Only you would know if they preprocessing filters you tried will help your network classify the images better than with no preprocessing filters. I suggest you look at the images where the classification was wrong and try to figure out how to make those images look like the images your network worked well on.

We don't even know what your images look like since you forgot to attach any here. And you didn't say what you want to know. Are you trying to apply one class to the whole image (like whether or not it has puppies in it) or do you want to classify on a pixel-by-pixel basis (like indicate which pixels are puppy pixels and which are background pixels)?

Why do you think that preprocessing will help? Maybe you don't even need to do it. Does it fail badly if you don't preprocess?

Do you know what an ROC curve is? Maybe you should build one up to determine which filter is best. https://en.wikipedia.org/wiki/Receiver_operating_characteristic MATLAB has ROC functions roc, rocmetrics, etc. that might help.

[Dina]

• Sample of images are attached, the images are ultrasound images
• The transfer learing for the network used for binary classification as normal and abnormal.
• i used filters like quassian,median,wiener ,... filters that might be useful to enhance the quality of the ultrasound images.
• i have a folder of the original noisy image and a folder for each applied filter

i dont know if the filtering is useful in training the network or not , and if it will affect in the training of the network to high the validation accuracy , which better in that case to train the network with each specific filtered images and etc or to calculate factors or values to know which one of these filters is the best without training the full network for each filtered images folder @Image Analyst

[Dina]

@Image Analyst

Answer 2

[Jaimin]

Hi @Dina,

I understand that you want to evaluate existing pre-trained models using images processed with different types of filters to determine which approach most suitable for binary classification during transfer learning.

I can suggest you a 2-step process to achieve this.

First step: Feature extraction

Evaluate images with different filters using a pre-trained model (e.g., "resnet50") and fetch the feature vector from that model, follow these steps:

1. Load the Pre-trained Model
2. Prepare Image Groups
3. Extract Features

Second step: Analyse Features

Collect the feature vectors for all images in each group (e.g. normal and abnormal). These vectors can then be used to perform statistical analysis, such as Principal Component Analysis (PCA) or calculating silhouette scores, to evaluate the separability and quality of the features extracted from each filtered group.

Compare the results from the different filtered groups to determine which filter produces the most discriminative features. This helps you identify the best filtering method to use before starting the transfer learning process.

For your reference, I have attached some useful resources

Performing transfer learning using resnet50: https://www.mathworks.com/help/deeplearning/ref/resnet50.html

Principal Component Analysis: https://www.mathworks.com/help/stats/pca.html

Silhouette plot: https://www.mathworks.com/help/stats/silhouette.html

Other pretrained models: https://www.mathworks.com/help/deeplearning/ug/pretrained-convolutional-neural-networks.html

I hope this will be helpful to you.