Infrared Small Target Detection (IRSTD) aims to detect small and dim targets in complex backgrounds. However, the low signal-to-noise ratio and reduced contrast in the infrared domain make it challenging to extract these targets, as the cluttered background can easily overpower them. Existing Convolutional Neural Networks (CNN)-based methods for IRSTD often suffer from information loss due to inadequate utilization of acquired information after downsampling operations. This limits their ability to accurately extract shape information related to infrared small targets. To address this challenge, we propose a Multi-scale U-shape Pyramid Transformer Network (MUPT-Net). Our network incorporates the U-shape Interaction Module (UIM) and the Multi-scale ViT Module (MSVM) to perform feature extraction. By fully leveraging and integrating the information obtained after each downsampling operation, our approach enables precise extraction of shape information for infrared small targets. Additionally, we introduce the Axial Compression Attention module (ACA), which focuses on capturing the interplay of positional information within the feature map to facilitate accurate detection of small targets. Through iteratively fusing and augmenting multi-scale features, our MUPT-Net effectively assimilates and harnesses contextual information of small targets. Experimental results on the SIRST v1, SIRST v2 and NUDT-SIRST datasets demonstrate the superiority of our approach compared to representative state-of-the-art (SOTA) IRSTD methods.

中文翻译：

---

MUPT-Net：用于红外小目标检测的多尺度U型金字塔变压器网络

红外小目标检测（IRSTD）旨在检测复杂背景下的小而昏暗的目标。然而，红外域的低信噪比和降低的对比度使得提取这些目标变得具有挑战性，因为杂乱的背景很容易压倒它们。现有的基于卷积神经网络 (CNN) 的 IRSTD 方法经常会由于下采样操作后获取的信息利用不足而导致信息丢失。这限制了他们准确提取与红外小目标相关的形状信息的能力。为了应对这一挑战，我们提出了多尺度 U 形金字塔变压器网络（MUPT-Net）。我们的网络结合了 U 形交互模块（UIM）和多尺度 ViT 模块（MSVM）来执行特征提取。通过充分利用和整合每次下采样操作后获得的信息，我们的方法能够精确提取红外小目标的形状信息。此外，我们还引入了轴向压缩注意力模块（ACA），该模块专注于捕获特征图中位置信息的相互作用，以促进小目标的准确检测。通过迭代地融合和增强多尺度特征，我们的 MUPT-Net 有效地同化和利用小目标的上下文信息。 SIRST v1、SIRST v2 和 NUDT-SIRST 数据集上的实验结果表明，与代表性的最先进 (SOTA) IRSTD 方法相比，我们的方法具有优越性。