High-resolution processing and sigmoid fusion modules for efficient detection of small objects in an embedded system.

Area of Science:

• Computer Vision
• Deep Learning
• Embedded Systems

Background:

• Deep learning enables accurate object detection for diverse applications, including autonomous vehicles and medical imaging.
• Small object detection is crucial for embedded systems in areas like drone reconnaissance and microscopic analysis.
• Existing methods often face challenges with computational cost and accuracy for small objects.

Purpose of the Study:

• To develop a lightweight and efficient small object detection model for embedded systems.
• To improve the accuracy and robustness of detecting small objects in challenging environments.
• To reduce the computational resources required for high-performance object detection.

Main Methods:

• Proposed a novel light-weight small object detection model incorporating a high-resolution processing module (HRPM) and a sigmoid fusion module (SFM).
• HRPM efficiently learns multi-scale features of small objects with reduced computational cost.
• SFM mitigates mis-classification errors by adjusting weights on lost small object information.

Main Results:

• The combined HRPM and SFM significantly improved detection accuracy with low computational overhead.
• The proposed model achieved higher mean average precision (mAP) on two-times higher-resolution input images compared to the original YOLOX-s.
• It utilized 57% fewer model parameters and 71% less computation (Gflops) than the baseline.

Conclusions:

• The developed model offers a significant improvement in small object detection accuracy and efficiency for embedded systems.
• It demonstrates strong performance in real-world scenarios, such as detecting small vehicles in drone reconnaissance imagery.
• This approach provides a viable solution for resource-constrained applications requiring precise small object detection.