CenterADNet: Infrared Video Target Detection Based on Central Point Regression

Area of Science:

• Computer Vision
• Infrared Imaging
• Deep Learning

Background:

• Infrared video target detection is crucial for warning and tracking systems.
• Weak targets in long-distance infrared images are challenging due to low contrast and background noise.
• Existing deep learning methods struggle with small target feature loss caused by downsampling.

Purpose of the Study:

• To develop a novel infrared video weak-target detection network.
• To address the limitations of current deep learning approaches in detecting small, low-visibility targets.
• To improve detection accuracy by suppressing background interference and preserving target features.

Main Methods:

• A new network based on central point regression is proposed.
• Feature fusion between consecutive and original frames suppresses background noise.
• High-resolution feature preservation and a spatial-temporal attention module are employed to capture target details.

Main Results:

• The proposed method demonstrates superior performance on a weak aircraft target detection dataset.
• Effective detection was also achieved on a simulated dataset based on real-world observations.
• The approach shows significant efficiency in detecting weak point targets in infrared continuous images.

Conclusions:

• The developed network effectively detects weak point targets in infrared video.
• The combination of feature fusion, high-resolution preservation, and attention mechanisms enhances detection accuracy.
• This method offers an efficient solution for critical infrared surveillance applications.