Ivan-ISTD: Rethinking Cross-domain Heteroscedastic Noise Perturbations in Infrared Small Target Detection

TL;DR

This work tackles cross-domain distribution shifts and heteroscedastic noise in infrared small target detection by introducing Ivan-ISTD, a two-stage framework combining Wavelet-guided Cross-domain Synthesis for data-space alignment with Real-domain Noise Invariance Learning for feature-space robustness. A Dynamic-ISTD Benchmark is proposed to simulate real-world degradations and evaluate cross-domain performance; extensive experiments demonstrate state-of-the-art results under both o.o.d and i.i.d conditions across multiple datasets. The approach leverages wavelet-based background separation, SSIM-guided target re-generation, and a noise-informed dual-branch network to build a dynamic noise library and enforce invariance. The work offers practical improvements for robust UAV infrared sensing and provides a public codebase to facilitate adoption and further research.

Abstract

In the multimedia domain, Infrared Small Target Detection (ISTD) plays a important role in drone-based multi-modality sensing. To address the dual challenges of cross-domain shift and heteroscedastic noise perturbations in ISTD, we propose a doubly wavelet-guided Invariance learning framework(Ivan-ISTD). In the first stage, we generate training samples aligned with the target domain using Wavelet-guided Cross-domain Synthesis. This wavelet-guided alignment machine accurately separates the target background through multi-frequency wavelet filtering. In the second stage, we introduce Real-domain Noise Invariance Learning, which extracts real noise characteristics from the target domain to build a dynamic noise library. The model learns noise invariance through self-supervised loss, thereby overcoming the limitations of distribution bias in traditional artificial noise modeling. Finally, we create the Dynamic-ISTD Benchmark, a cross-domain dynamic degradation dataset that simulates the distribution shifts encountered in real-world applications. Additionally, we validate the versatility of our method using other real-world datasets. Experimental results demonstrate that our approach outperforms existing state-of-the-art methods in terms of many quantitative metrics. In particular, Ivan-ISTD demonstrates excellent robustness in cross-domain scenarios. The code for this work can be found at: https://github.com/nanjin1/Ivan-ISTD.

Figures (9)

• Figure 1: A significant distribution shift between the source and target domains. Our transfer background method helps mitigate this domain shift issue.
• Figure 2: Real domain noise (green) and artificially synthesized noise (red). The real domain noise exhibits greater variability, whereas the synthetic noise is uniform and cannot capture the dynamic heteroscedastic nature of noise in real-world situations.
• Figure 3: The overall framework of our method. Stage I presents the complete Wavelet-guided Cross-domain Synthesis framework, consisting of (a) and (b). (a) illustrates the process of extracting high-value small targets from the source set (right image) and unsupervised background sample extraction (left image). (b) describes the synthesis of new samples using SSIM ranking and Poisson blending techniques. Stage II outlines the process of Real-domain Noise Invariance Learning. It involves unsupervised real noise sampling from target images, minimizing loss to enhance the model's robustness to cross-domain noise. (c) shows the detailed process of extracting noise from the target set.
• Figure 4: (a) illustrates the proposed unsupervised Background Region Detection framework. After applying Wavelet Multi-frequency Filtering (WMF) and calculating small target probabilities based on edge density and Laplacian operator, we select appropriate background samples. (b) shows the detailed implementation of WMF, where $L^{low}$ is the low-frequency sub-band obtained through wavelet decomposition, and $\{H_k^{high}\}^3_{k=1}$ is the high-frequency sub-band.
• Figure 5: Visualization results of Background Region Detection methods on target data. The left column displays the background region tendency map, while the right column presents the cropped region. The red boxes indicate the cropping results from different methods, and the yellow dashed circles highlight cases where other methods misclassified the background, including unnecessary small targets within the background. The yellow boxes are used to clearly emphasize the small targets.