Is my patternnet for two-class image classification okay?

I have 10,000 images for two classes including 5000 for each class. I used all default parameters. One-fifth of the whole is set as holdout set and for the training network, 70:15:15 ratio is used for training, testing, and validation set inside the network. So, And then, I trained my network with many hidden layers. Among them, 340 hidden layer size has highest accuracy for confusion matrix as I have attached.And, the accuracy for the holdout set is 92.45%. And the performance of the network is as follows:

I would like to know whether my network is okay or not. And I also want to know which more parameters I should vary to get better accuracy.

.

Answer 1

[Kenta]

Dear San May

I think your work is great! Cross entropy error of training data is much lower than that of test data.

So, maybe your network is over-fitted to the training data. To prevent it, it can be considered to change L2 regularization, momentum, dropout and so on. Depending on your optimizar (e.g. sgdm, adam...), other parameters can be corrected.

Please refer the webpage below about the parameter setting.

https://jp.mathworks.com/help/deeplearning/ref/trainingoptions.html?lang=en

In my opinion and experience, it would be better to use transfer learning.

It utilizes pre-trained network and it is better than to construct your network from scratch.

https://jp.mathworks.com/help/deeplearning/examples/transfer-learning-using-alexnet.html?searchHighlight=alexnet&s_tid=doc_srchtitle

We can easily try that with the document. In my case, the accuracy was much better to use the pre-trained network such as Alexnet and GoogleNet.

Answer 2

[San May]

Yes, thank you so much.

Answer 3

[Dhia El Hak Daamouche]

Hey bro,

I'm doing the same work (I wanna classify an image of Dubai into two classes: urban>> Black and non_urban>> white). I've completed the training and had the results as you. Now what I want is to classify the image by using this training, but I couldn't figure out how to do it. do you have an idea how to do the classification?

Any help would be so approciate

regards,

Dhia