ISTD-diff: Infrared Small Target Detection via Conditional Diffusion Models

• Published in: IEEE geoscience and remote sensing letters Vol. 21; pp. 1 - 5
• Main Authors: Du, Nini, Gong, Xuemei, Liu, Ye
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Computer vision; Correlation; Diffusion model; Diffusion models; dual-dimension transformer; feature injection; Infrared imagery; infrared small-object detection; Noise measurement; Noise reduction; Object detection; Rivers; Streaming media; Target detection; Target masking; Task analysis; Transformers
• ISSN: 1545-598X; 1558-0571
• DOI: 10.1109/LGRS.2024.3401838

Infrared small-target detection (IRSTD), which is to extract tiny and dim targets that are hidden in noisy and messy backgrounds, is a challenging task in computer vision. Inspired by the recently powerful deep generative models, we formulate the IRSTD as a generative task and design a conditional denoising (DE) model termed ISTD-diff to iteratively generate the target mask from the noisy one. In addition, ISTD-diff employs a two-pathway architecture, consisting of a conditional prior (CP) stream for encoding the input infrared image prior and a DE stream for cleaning up the noisy masks. Both streams are equipped with several cascaded innovative channel-dimension transformer (CDT) layers, which capture the global correlations efficiently and reduce computational demands effectively. Moreover, to strengthen the DE learning process, we proposed a simple, but powerful method named attention injection module (AIM), which provides detailed control over the DE stream. Extensive experiments finely demonstrate the superior performance of our ISTD-diff beyond the current representative segmentation-based state-of-the-art (SOTA) algorithms.