MEAC: A Multi-Scale Edge-Aware Convolution Module for Robust Infrared Small-Target Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Infrared small-target detection is crucial for military, environmental, and safety applications.
• Traditional Convolutional Neural Networks (CNNs) face challenges with small, low-contrast targets due to limited feature extraction.
• Complex backgrounds and low signal-to-noise ratios further complicate detection.

Purpose of the Study:

• To introduce a novel Multi-Scale Edge-Aware Convolution (MEAC) module for enhanced infrared small-target detection.
• To improve feature representation without increasing computational cost or parameter count.
• To address the limitations of existing CNNs in detecting weak and small infrared targets.

Main Methods:

• Proposed the Multi-Scale Edge-Aware Convolution (MEAC) module.
• MEAC fuses local features, multi-scale context via dilated convolutions, and edge cues from differential Gaussian filters.
• Integrated channel and spatial attention mechanisms for adaptive feature emphasis.

Main Results:

• Networks augmented with MEAC significantly outperformed baseline models on three public datasets (SIRSTD-UAVB, IRSTDv1, IRSTD-1K).
• MEAC achieved superior detection accuracy and robustness compared to eleven mainstream convolutional modules.
• Consistent performance improvements were observed across YOLOv10, YOLOv11, and YOLOv12 architectures.

Conclusions:

• The MEAC module effectively enhances the detection of small, weak infrared targets.
• MEAC demonstrates strong generalization ability and practical application potential in complex scenarios.
• The proposed module offers a significant advantage in suppressing background noise and improving detection accuracy.