Data Augmentation in Classification and Segmentation: A Survey and New Strategies

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks, especially convolutional neural networks, have transformed computer vision.
• A significant challenge in deep learning is the need for large datasets, as data scarcity can lead to overfitting.
• Data augmentation is a key strategy to address the issue of insufficient training data.

Purpose of the Study:

• To survey existing data augmentation techniques in computer vision.
• To propose novel strategies for data augmentation, focusing on the use of local image information.
• To introduce a new parameter-free and easily implementable data augmentation method: the random local rotation strategy.

Main Methods:

• Surveying current data augmentation methods for computer vision tasks such as segmentation and classification.
• Developing a novel data augmentation technique involving the random selection of circular regions within an image and rotating them at random angles.
• Implementing the random local rotation strategy as an alternative to traditional rotation methods.

Main Results:

• The proposed random local rotation strategy effectively addresses limitations of traditional rotation techniques, such as irregular image boundaries.
• Experimental results demonstrate that the new strategy consistently outperforms traditional counterparts in image classification tasks.
• The random local rotation strategy can be used independently or in conjunction with other data augmentation methods.

Conclusions:

• The random local rotation strategy offers a robust and efficient solution for data augmentation in computer vision.
• This technique helps mitigate overfitting issues caused by limited datasets.
• The proposed method provides a valuable alternative and complement to existing data augmentation approaches, enhancing model performance.