Augmentation of Alzheimer Images Base on Visual Transform
Authors:
- 10 January 2023
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 176-180
Abstract
With the continuous development of deep learning, transformer has become a new leader in the field of vision in recent years. Transformer has not only completely changed the field of natural language processing (NLP), but also made some pioneering work in the field of computer vision (CV). Compared with convolutional neural network (CNN), Visual Transformer (ViT) has achieved excellent performance on many benchmarks such as ImageNet, COCO and ADE20k, depending on its excellent modeling ability. In the process of experiments, the final results are often overfitted due to insufficient training data, too many iterations and too complex models. In view of the over fitting caused by insufficient training data, the most direct way is to collect more training data. However, data acquisition is often the most difficult part. In the field of CV, we can enrich the distribution of training data by using data augmentation, so that the model obtained through the training set has stronger generalization ability. In this paper, from the perspective of training data processing, Alzheimer's disease data set is taken as a sample, and the sample data set is enhanced by Flip and rotation, Cutmix and Mixup. The results show that the accuracy in test data of the base model is 86.51%. Then, the Flip and rotation augmentation reaches 86.79%, a lower one, shows this data augmentation is not suitable for this task. The Cutmix reaches 88.95%, which improves the prediction accuracy of the model. The effect of Mixup is the best, reaching 89.61%. The Mixup augmentation method can effectively improve accuracy and overcome overfitting.Keywords
This publication has 3 references indexed in Scilit:
- Facial Expression Classification Based on Ensemble Convolutional Neural NetworkLaser & Optoelectronics Progress, 2020
- CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features2019
- mixup: Beyond Empirical Risk MinimizationarXiv, 2018