Abstract: Aiming at the challenges of single-frame infrared small target detection, we proposed an infrared small target detection method based on feature separation and global context. Firstly, aiming at the problem of insufficient features of small targets, we designed a feature separation module that captured the difference of target background contrast  by central differentical convolution, and combined  fast Fourier convolution to extract the edge gradient information, achieving  efficient separation of target features and background noise. Secondly, aiming at the problem of feature loss caused by downsampling, we constructed a global context extraction module  to perform cross-scale global modeling of deep features,  preventing the loss of target features in the deep layers of the network. The experimental results  on multiple public datasets show that this method  significantly improves mIoU, nIoU and F₁ indicators compared with advanced algorithms such as AGPCNet and DNANet,  optimizes the performance of infrared small target detection algorithm and improves the perception ability of complex scenes.

Key words: small target detection, infrared image, central differential convolution, feature adaptation, deep learning