核心概念
The author proposes a diffusion model framework for Infrared Small Target Detection to address target-level insensitivity by generating mask posterior distributions.
摘要
The content discusses the challenges of detecting small infrared targets due to low signal-to-clutter ratio and proposes a generative approach using diffusion models. The method compensates for false alarms and missed detections, achieving competitive performance on various datasets. The Low-frequency Isolation module in the wavelet domain reduces interference from infrared noise, enhancing detection accuracy.
Key points include the introduction of Infrared Small Target Detection (IRSTD), challenges faced by existing methods, proposal of a diffusion model framework, design of a Low-frequency Isolation module, contributions of the proposed method, related work in IRSTD, and comparison with state-of-the-art methods.
The experiments conducted on three datasets demonstrate the effectiveness of the proposed IRSTD-Diff method in improving target-level sensitivity and reducing false alarms. The ablation studies confirm the importance of both Conditional Encoder (CE) and Low-frequency Isolation in the Wavelet domain (LIW) modules in enhancing performance.
統計資料
Experiments show that IRSTD-Diff achieves competitive performance gains over state-of-the-art methods on NUAA-SIRST, IRSTD-1k, and NUDT-SIRST datasets.
False alarm rate typically falls within the order of 10^-6 for prevailing state-of-the-art methods.
Our IRSTD-Diff demonstrates superior performance at the target-level compared to other discriminative methods.
The best k value for LIW is 2 for 256 resolution.
Training steps can be reduced to smaller values like 30 to enhance efficiency without compromising performance.
引述
"The final detection results are obtained by sampling from this distribution."
"Conventional unlearnable methods are hindered by numerous hyper-parameters."
"Our approach compensates for this issue from the perspective of generative learning."