Diffeomorphic transforms for data augmentation of highly variable shape and texture objects.

Area of Science:

• Computer Vision
• Machine Learning
• Bioinformatics

Background:

• Deep convolutional neural networks (CNNs) require large, labeled datasets for training, which are often unavailable in specialized fields like biology and medicine.
• Existing data augmentation methods can introduce artifacts or unrealistic spatial variations, limiting their effectiveness for training deep learning models.
• Expert knowledge is crucial for accurate image classification, particularly when subtle features or significant variations within a class exist.

Purpose of the Study:

• To develop a novel data augmentation procedure for generating realistic image samples.
• To address the challenge of limited labeled data in specialized domains by creating diverse training sets.
• To improve the accuracy of deep learning models in image classification tasks.

Main Methods:

• A new data augmentation technique combines two samples from the same class to generate novel, realistic images.
• The procedure utilizes morphing and image registration methods, specifically employing diffeomorphic transformations.
• The method was tested on diatom, glomeruli, and pollen identification problems.

Main Results:

• The proposed technique demonstrated accuracy improvements of 0.47% for diatoms, 1.47% for glomeruli, and 0.23% for pollen.
• The method successfully simulated shape changes in diatom life cycles and added diverse shapes/textures for glomeruli.
• Performance gains were observed even in the pollen dataset, which has smaller inter-class variations and noise.

Conclusions:

• The novel data augmentation method effectively generates realistic intermediate shapes, mimicking natural variations in biological samples.
• This approach is particularly beneficial for datasets with significant shape and texture diversity, such as glomeruli.
• The technique offers a valuable alternative to standard data augmentation, outperforming other methods in specific challenging scenarios.