基于双流学习框架的红外小目标检测研究

Research on Infrared Small Target Detection based on Dual-stream Learning Framework

为解决由于目标尺寸小、背景复杂等因素导致红外小目标检测精度难以提高的问题,提出一种基于双流学习框架红外小目标检测方法。将分割网络用于小目标检测,并将超分辨率任务作为辅助手段,引入共享特征注意力机制(SFAM),解决特征融合和迭代中的特征损失问题。通过在公共数据集上进行了4种不同场景的广泛实验,并以0.835的精度优于其他方法。同时,消融研究也证实了SFAM的重要性和可行性。

In order to solve the problem that the accuracy of infrared small target detection is difficult to improve due to the factors such as small target size and complex background,a method of infrared small target detection based on dual-stream learning framework was proposed.The segmented network was used for small target detection,and the super-resolution task was used as an auxiliary means,the shared feature attention mechanism(SFAM)was introduced to solve the feature loss problem in feature fusion and iteration.By conducting extensive experiments on 4 different scenarios on a public dataset,the proposed method scores better than other methods with an accuracy of 0.835.At the same time,the ablation study also confirmed the importance and feasibility of SFAM.