It GAN Do Better: GAN-Based Detection of Objects on Images With Varying Quality

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection models often struggle with reduced image quality, impacting performance.
• Image quality degradation affects both classification and bounding box regression accuracy.
• Existing methods lack robustness for real-world, variable image quality scenarios.

Purpose of the Study:

• To propose a novel Generative Adversarial Network (GAN)-based framework, GAN-DO, for robust object detection.
• To enhance deep neural network (DNN) architectures for improved performance on low-quality images.
• To evaluate the impact of image quality on object detection tasks.

Main Methods:

• Developed a GAN-based Detection of Objects (GAN-DO) framework adaptable to various DNN architectures.
• Generated features using GANs to improve robustness against reduced image quality.
• Evaluated the framework's performance on object detection datasets using state-of-the-art baseline models.

Main Results:

• The GAN-DO framework demonstrated improved robustness to varying image quality.
• Achieved higher mean Average Precision (mAP) compared to existing approaches.
• The framework maintained baseline model architecture, complexity, and inference speed.

Conclusions:

• The proposed GAN-DO framework effectively enhances object detection robustness for reduced-quality images.
• GAN-DO offers a generalizable solution applicable to multiple DNN architectures.
• The method provides a significant improvement in object detection accuracy under challenging image conditions.