Random walk-based erasing data augmentation for deep learning

Zhang, Chao; Zhong, Weifeng; Li, Changfeng; Deng, Haipeng

doi:10.1007/s11760-022-02461-3

Cited by 1 publication

(1 citation statement)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Data augmentation-which tends to manipulate the datahas been proposed as the second solution to solve the problem of imbalanced data and produce synthetic data [29]. Augmenting data, many researchers use methods such as rotating the input image, blurring, sharpening, and cutting the image [30]. Being in the input space, in spite of proper response, these methods take a long time.…”

Section: Introductionmentioning

confidence: 99%

A Multi-level Deep Neural Network to Diagnose Coronavirus Disease with Imbalanced Data

Zabbah,

Layeghi,

Ebrahimpour

2024

Preprint

View full text Add to dashboard Cite

One of the main challenges of medical data mining is the classification of imbalanced datasets. Often the data which are used for training classification is of no proper distribution. It occurs when a class has few samples in nature inclining the learner model to the major class. This imbalance is quite obvious in the data X-ray image of coronavirous, due to its newness, so the samples of healthy and pneumonia are more than COVID- 19 ones. Here we proposed a multi-level model that improves diagnosis of the disease using data augmentation in the minor class of coronavirous cases. The model contains a deep feature extractor, a novel algorithem to find the scatter topolgy of the minor class, a mechanism to selectively generate synthetic samples and finally clasify input data. Coronavirous images are fed to a deep neural networks as a feature estractor. Then, finding scatter topology of the minor class, synthetic samples are generated in the feature space and evaluated by some expertes so that the synthetic samples resemble real ones mostly. Finally, the pool of synthetic and real samples of each classes are transferred to a discriminator network after the dataset is balanced. Results indicate that the production of samples by the framework not only improves classification performance but also has a more desirable performance than other data augmentation methods.

show abstract

Section: Introductionmentioning

confidence: 99%