Robust Data Augmentation Generative Adversarial Network for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection models often require large datasets for optimal performance.
• Small datasets limit the accuracy and generalizability of object detection models.
• Generative Adversarial Networks (GANs) show promise for data augmentation but require careful application.

Purpose of the Study:

• To propose a novel GAN-based data augmentation method, Robust Data Augmentation GAN (RDAGAN), specifically for small object detection datasets.
• To improve the performance of object detection models, such as YOLOv5, through effective data augmentation.
• To address limitations in existing GAN-based augmentation by preserving background information and localizing object generation.

Main Methods:

• RDAGAN employs a pipelined approach with two distinct networks: an object generation network and an image translation network.
• The object generation network creates object instances based on bounding boxes from the input dataset.
• The image translation network seamlessly integrates these generated objects into clean background images.

Main Results:

• Quantitative experiments demonstrated significant improvements in YOLOv5 fire detection performance using RDAGAN-generated data.
• Comparative evaluations confirmed RDAGAN's ability to preserve background context and accurately localize object generation.
• Ablation studies validated the critical contribution of each component within the RDAGAN framework.

Conclusions:

• RDAGAN effectively augments small datasets for object detection tasks, leading to enhanced model performance.
• The proposed method offers a robust solution for data scarcity issues in object detection.
• RDAGAN's architecture provides fine-grained control over object generation and integration, outperforming generic augmentation techniques.