Using Transfer Learning Approach for Down Syndrome Features Extraction and Data Augmentation for Data Expansion

Abstract

Purpose: People with Down Syndrome must be served specially because they have an intellectual disability with abnormality in memory and learning, so creating a model for DS recognition may provide safe services to them, using the transfer learning technique can improve high metrics with a small dataset, depending on previous knowledge, there is no available Down syndrome dataset, one can use to train.

Materials and Methods: A new dataset is created by gathering images, two classes (Down=209 images, non-Down=214 images), and then expanding this dataset using Augmentation to create the final dataset of 892 images (Down=415 images, Non-Down=477 images). Finally, using a suitable training model, in this work, Xception and Resnet models are used, and the pre-trained models are trained on Imagenet dataset, which consists of (1000) classes.

Results: By using the Xception model and the Resnet model, it is concluded that when using the Resnet model the accuracy is 95.93% and the loss function is 0.16, while by using the Xception model, the accuracy is 96.57% and the loss function is 0.12.

Conclusion: Transfer learning is used to overcome the suitability of dataset size and minimize the cost of training, and time processing the accuracy and loss function is good when using the Xception model, in addition, the Xception metrics are the best compared with the previous studies.