Single-frame InfraRed Small Target-V2

This dataset is from "One-Stage Cascade Refinement Networks for Infrared Small Target Detection." It includes 427 infrared images and 480 targets (due to the lack of infrared sequences, SIRST also contains infrared images at a wavelength of 950 nm, in addition to shortwave and midwave infrared images). Approximately 90% of the images contain only one target, while about 10% have multiple targets (which may be overlooked in sparse/significant methods due to global unique assumptions). Around 55% of the target area occupies less than 0.02% (i.e., in a 300x300 image, the target pixel size is 3x3). Only 35% of the targets have the highest brightness in the entire image. Considering that 65% of the targets have brightness levels very similar to or even darker than the background, we cautiously employ the target saliency assumption. We use this dataset for research on infrared small targets and propose a lightweight detection algorithm for single-frame infrared small targets.

本数据集源自论文《单级级联细化网络用于红外小目标检测》（One-Stage Cascade Refinement Networks for Infrared Small Target Detection）。该数据集包含427幅红外图像与480个目标。由于红外序列数据不足，SIRST数据集除收录短波红外与中波红外图像外，还包含波长为950nm的红外图像。约90%的图像仅包含单个目标，剩余约10%的图像含有多个目标，这类多目标场景易因多数稀疏/显著性检测方法基于全局唯一性假设而被遗漏。约55%的目标占图像总面积不足0.02%——以300×300分辨率的图像为例，对应目标的像素尺寸仅为3×3。仅35%的目标为整幅图像中亮度最高的对象；鉴于65%的目标亮度与背景极为接近，甚至低于背景亮度，因此我们谨慎采用目标显著性假设开展相关研究。本研究使用该数据集开展红外小目标检测相关研究，并提出了一种针对单帧红外小目标的轻量化检测算法。

搜集汇总

背景与挑战

背景概述

Single-frame InfraRed Small Target-V2数据集包含427张红外图像和480个目标，主要用于红外小目标检测研究。数据集特点包括目标数量、面积和亮度分布的多样性，约90%的图像仅含一个目标，55%的目标面积小于0.02%，且仅35%的目标具有最高亮度。