MSIA-Net: A Lightweight Infrared Target Detection Network with Efficient Information Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Infrared target detection faces challenges with large models and numerous parameters.
• Existing methods often struggle with information loss during down-sampling and noise during feature fusion.

Purpose of the Study:

• To propose a lightweight and efficient infrared target detection network, MSIA-Net.
• To reduce model complexity while enhancing detection accuracy and target focus.

Main Methods:

• Introduced MSIA (Multi-Scale Asymmetric) module for efficient feature extraction using asymmetric convolution.
• Developed DPP (Down-sampling with Pooling) module to minimize information loss during down-sampling.
• Proposed LIR-FPN (Lightweight Information-Recycling Feature Pyramid Network) for effective feature fusion with reduced noise.
• Integrated Coordinate Attention (CA) into LIR-FPN to improve target localization and feature representation.

Main Results:

• MSIA-Net demonstrated significant reduction in model parameters compared to existing methods.
• The proposed modules effectively reduced information loss and noise during processing.
• Integration of CA enhanced the network's ability to focus on target features.
• Comparative experiments on the FLIR on-board infrared image dataset validated superior detection performance.

Conclusions:

• MSIA-Net offers a powerful and efficient solution for infrared target detection.
• The novel architectural components contribute to improved performance and reduced computational cost.
• The network shows strong potential for real-world infrared imaging applications.