Centered Multi-Task Generative Adversarial Network for Small Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks have advanced object detection but struggle with small objects.
• Gaze estimation technology has significantly improved with visual sensor development.
• Integrating object detection and gaze estimation offers potential for enhanced small object detection.

Purpose of the Study:

• To present a Centered Multi-Task Generative Adversarial Network (CMTGAN) for improved small object detection.
• To combine image super-resolution and two-stage small object detection within a single GAN framework.
• To enhance the performance of detecting small objects by leveraging gaze estimation.

Main Methods:

• Developed a GAN with a generator for image super-resolution and a discriminator for object detection.
• Introduced an artificial texture loss in the generator to preserve small object features.
• Utilized a centered mask in the generator to focus on image regions likely to contain small objects.
• Proposed a discriminator with detection loss for two-stage small object detection, adaptable to other GANs.

Main Results:

• CMTGAN generates explicit super-resolution images with richer information compared to interpolation methods.
• The proposed method demonstrates superior small object detection performance over mainstream approaches.
• Experiments validate the effectiveness of the artificial texture loss and centered mask.

Conclusions:

• CMTGAN effectively integrates super-resolution and object detection for improved small object recognition.
• The novel approach offers a significant advancement in addressing the challenges of small object detection.
• The proposed discriminator architecture is a valuable contribution to GAN-based object detection.