Progressive Fusion of Multi-Scale Mamba Context and Local Detail Priors for Infrared Small Target Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Infrared Small Target Detection (IRSTD) relies on modeling long-range dependencies, leading to the adoption of Transformer architectures.
• Transformers offer improved global context but suffer from high computational costs, limiting practical applications.
• Mamba presents an efficient alternative for long-range dependencies, yet lacks specific mechanisms for fine-grained local details crucial for IRSTD.

Purpose of the Study:

• To propose MCFNet, an encoder-decoder framework designed to enhance Infrared Small Target Detection (IRSTD) performance.
• To address the limitations of Mamba in capturing local details and multi-scale background variations for IRSTD.
• To achieve high target-level detection accuracy with reduced computational complexity.

Main Methods:

• Developed MCFNet, an encoder-decoder framework integrating Mamba for efficient long-range dependency modeling.
• Introduced a Detail-Capturable Convolution Block to enhance local feature perception.
• Incorporated a Multi-scale Contextual Mamba Block for improved multi-scale background modeling.
• Designed a Feature Fusion Decoding Module to effectively combine global and local representations.

Main Results:

• MCFNet demonstrated superior performance on multiple public IRSTD benchmark datasets.
• The proposed model achieved higher detection accuracy compared to existing methods.
• MCFNet significantly reduced false alarms in Infrared Small Target Detection.
• The model offers improved target-level detection with moderate computational cost.

Conclusions:

• MCFNet effectively addresses the limitations of Mamba for IRSTD by incorporating specialized blocks for local details and multi-scale context.
• The proposed framework achieves state-of-the-art performance in Infrared Small Target Detection, balancing accuracy and efficiency.
• MCFNet provides a promising solution for practical IRSTD applications demanding high accuracy and low false alarm rates.