I am working on a novel image classification problem with only 128 images of one class and 25 images of another class. From CNN perspective, I have too little data to learn features well and on top of that, the dataset is highly imbalanced. What strategies should I adopt to improve the model in this scenario? Could other techniques could be tried (SVM ?)
Dear Ashutosh Karna
I think this is an interesting question.
Since deep learning required a large dataset, if we have a limited dataset, transfer learning with freeze, hyperparameter tuning is one approach. I have a little bit of work about limited pathological datasets. Please site :
Article Conditional Sliding Windows: An Approach for Handling Data L...
Toto Haryanto Thank you for sharing the paper. I would check the methodology in detail.
Ashutosh Karna , I think you already know , why we used CNN for image Classification problems. If your dataset size is large , then simple 2D CNN & Transfer learning models will be ok.
otherwise, you can try SVM & others ml classification models..but model won't perform well if data set is too small .
Here is the sample code: https://www.kaggle.com/ashutoshvarma/image-classification-using-svm-92-accuracy
Check https://machinelearningmastery.com/cost-sensitive-neural-network-for-imbalanced-classification/
Hello Karna,
Kindly look at this paper.
Conference Paper Towards Effective Classification of Imbalanced Data with Con...
I think you can use Pre-Trained Models for Image Classification
these are the top 4 pre-trained models for image classification:
VGG-16
ResNet50
Inceptionv3
EfficientNet
Here is a table for you to refer to these models and their performance in the attached photo
The class imbalance is not the major issue here. You have too small dataset. In case the images are of ordinary kind, you may exploit Transfer Learning techniques. You may also try to employ data augmentation and the Few-Shot Classification approach. You may also try to expand you dataset using pre-trained GAN, however, this path is a tricky one and strongly relies on the characteristics of you data.
Mikhail Krinitskiy You are absolutely right. Class imbalance could still be dealt with but the lack of data is the biggest challenge. I haven't tried GAN yet to create synthetic images and not sure if it would help much but definitely, I can give it a try.
Ashutosh Karna Actually, GANs would be a last-line treatment. You need to be sure your data has the same characteristics (distribution of features) as the data a GAN was trained on. In this case, you may try to generate enough data to pre-train you network in a Self-Supervised manner (please refer to Contrastive Learning topic and something like DINO, SimCLRv2, etc.) One may also exploit huge datasets of unlabeled data (in case there are any) without GANs using these self-supervised approaches. Sometimes the networks trained withing these approaches beat supervisely-trained networks.
YOu can use also use data augmentation techniques to balanced and increase the number of images in the dataset.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification