OFIDA: Object-focused image data augmentation with attention-driven graph convolutional networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image data augmentation (DA) is vital for increasing training data quantity and diversity.
• Existing DA methods like image manipulation and generative models can distort images or fail to preserve spatial details.
• Challenges include accurate object representation and maintaining fine details during augmentation.

Purpose of the Study:

• To introduce OFIDA (Object-Focused Image Data Augmentation), an algorithm designed to overcome limitations of current DA techniques.
• To enhance the authenticity of augmented data by preserving essential target regions and simulating real-world distributions.
• To improve object comprehension in real-world scenarios through context-aware augmentation.

Main Methods:

• OFIDA utilizes a graph-based structure and object detection for streamlined augmentation.
• It leverages graph properties (connectivity, hierarchy) to capture object essence and context.
• Introduces DynamicFocusNet, an object detection algorithm employing dynamic graph convolutions and attention mechanisms.

Main Results:

• OFIDA implements one-to-many enhancements, preserving target regions and simulating realistic data.
• DynamicFocusNet effectively merges graph convolutions and attention for flexible receptive field adjustment.
• Experimental results demonstrate OFIDA's superiority over state-of-the-art methods on six benchmark datasets.

Conclusions:

• OFIDA effectively addresses limitations in image data augmentation by focusing on object preservation and realism.
• The proposed DynamicFocusNet enhances object detection within the graph framework for improved augmentation.
• OFIDA offers a superior approach for generating diverse and high-quality augmented image data.