Mitigate Target-level Insensitivity of Infrared Small Target Detection via Posterior Distribution Modeling

TL;DR

A diffusion model framework for IRSTD that maximizes the posterior distribution of the small target mask to surmount the performance bottleneck associated with minimizing discriminative empirical risk and achieves competitive performance gains over state-of-the-art methods on NUAA-SIRST, NUDT-SIRST, and IRSTD-1 k datasets.

Abstract

Infrared Small Target Detection (IRSTD) aims to segment small targets from infrared clutter background. Existing methods mainly focus on discriminative approaches, i.e., a pixel-level front-background binary segmentation. Since infrared small targets are small and low signal-to-clutter ratio, empirical risk has few disturbances when a certain false alarm and missed detection exist, which seriously affect the further improvement of such methods. Motivated by the dense prediction generative methods, in this paper, we propose a diffusion model framework for Infrared Small Target Detection which compensates pixel-level discriminant with mask posterior distribution modeling. Furthermore, we design a Low-frequency Isolation in the wavelet domain to suppress the interference of intrinsic infrared noise on the diffusion noise estimation. This transition from the discriminative paradigm to generative one enables us to bypass the target-level insensitivity. Experiments show that the proposed method achieves competitive performance gains over state-of-the-art methods on NUAA-SIRST, IRSTD-1k, and NUDT-SIRST datasets. Code are available at https://github.com/Li-Haoqing/IRSTD-Diff.

Figures (13)

• Figure 1: Schematic comparison between (a) Existing Discriminative Methods and (b) Our Generative Method for Infrared Small Target Detection. Our approach obtains the posterior distribution of small target masks.
• Figure 2: Sketch map of the Back-propagation process. The arrows to the right represent the Forward-propagation, while the arrows to the left represent the Back-propagation. For ease of representation, the bias and activation layers are omitted. $w_i$ and $h_i$ are the parameter and output of the $i$-th layer, respectively. All these partial derivatives in the IRSTD networks are exceedingly diminutive, but false alarms exist.
• Figure 3: Overview of the proposed diffusion framework for Infrared Small Target Detection. Each colored rectangle represents a corresponding Block, and the rectangle's size reflects the latent space level of the Blocks. The small gray arrows are in the same direction as the computational graph, while the large arrows are indicators.
• Figure 4: Low-frequency Isolation in the Wavelet domain (LIW). L / H represent the visualization of high / low-frequency components in the wavelet domain. R is the visualization of the estimated residuals.
• Figure 5: One-level 2-D Haar Discrete Wavelet Transform (HDWT) haar applied to the infrared features. The left figures are Conditional Encoder output features, and the right figures are corresponding wavelet domain visualizations. The bottom left, bottom right, top left, and top right subfigures are low-frequency approximation and high-frequency horizontal, vertical, and diagonal components, respectively. These wavelet domain components are used as part of LIW.