Answers are grounded in this paper and cite their sources.
In response to the challenges of small target size, low signal-to-noise ratio, and serious background clutter interference in infrared small target detection, this paper proposes what is believed to be a new lightweight network architecture—CMANet, whose core innovation lies in the design of two key modules: target feature enhancement module (TFEM) and multi-attention Mamba (MMamba). The TFEM effectively captures the typical Gaussian distribution characteristics of infrared small targets through the cross-shaped convolution structure and significantly enhances the initial response of weak targets in the shallow layer of the network, alleviating the problem of feature flooding caused by the lack of texture and edge information of the targets. MMamba introduces a multiple attention mechanism of channel and space on the basis of the Mamba state space model, which strengthens the model’s ability to model long-range dependencies in key areas while maintaining linear computational complexity, and overcomes the limitation of the original Mamba in visual tasks with a single perception dimension. The experimental results show that CMANet achieves excellent performance on the three benchmark datasets of IRSTD-1k, NUDT-SIRST, and SIRST. Under the lightweight configuration of only 0.639 M parameters and 0.829 G FLOPs, it not only achieves the leading segmentation accuracy and an extremely low false alarm rate but also achieves the real-time inference speed of 180.9 FPS, which fully verifies the effectiveness and collaborative advantages of the TFEM and the MMamba module in improving detection accuracy and efficiency.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: