I am trying to build a classification algorithm having 28 classes. These classes consists of Logo of companies like adidas , Nike etc. I have very low dataset below than 100 images and greater than 70 images. I have trained CNN model but not got decent results . Accuracy is not good. I want to switch on pretrained model. I don't know which pretrained model should I use either VGG , ResNet etc. because I saw on Internet that you may use Pretrained model on similar dataset but I have logo of companies , I don't know that any pre trained model is trained on such similar type of objects or not. How should I choose pre trained model that will performs well on my dataset.

  • 1
    $\begingroup$ We can't really know - as is typical with these type of things, you probably just have to give it a try and see for yourself. $\endgroup$ – Juho Jan 16 at 19:40
  • $\begingroup$ @Juho Any Guess ? $\endgroup$ – Hamza Jan 16 at 19:59
  • $\begingroup$ No way to know without trying it. $\endgroup$ – D.W. Jan 16 at 20:29

There's no general answer. In this field, you'll just have to try multiple approaches and see how well they work. We usually won't know what will work best until we try it.

I can suggest three candidate approaches:

  1. Get more data.

  2. Use a pre-trained model and fine-tune it. You can use any state-of-the-art model, pick a version that was pre-trained on ImageNet, and then fine-tune it (either fixing the first n-1 layers and training only the last layer; or training all layers, but with a smaller step size).

  3. Use few-shot learning. This will probably require more sophisticated methods and more study.

Aggressive data augmentation might be useful.

  • $\begingroup$ When I increase the number of Convolutional layers , accuracy goes upward with 2 percent. I have increased up to 4 Hidden layers and got an accuracy of 45% max . What is your perception , should I increase more ? $\endgroup$ – Hamza Jan 17 at 9:21
  • $\begingroup$ @Hamza, sorry, this site isn't a good fit for what you seem to be looking for. The site is not built for interactive back-and-forth interactions. $\endgroup$ – D.W. Jan 17 at 19:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.