Abstract:Aiming at the problems that the feature information extraction of dim and small targets is difficult in the remote infrared target detection technology, and the strong noise interference leads to the high false alarm rate and missing detection rate of the detection algorithm, a single-frame infrared dim and small target detection algorithm based on convolutional neural network(CNN)-Transformer is proposed. Aiming at the problem that the receptive field is limited and vulnerable to local noise interference when extracting the feature information of infrared dim and small targets with the existing methods, the window self-attention calculation module in Swin Transformer is improved, and a local perception enhancement module based on separable convolution is designed, which takes into account the extraction of global information and local information, and improves the ability of backbone network to extract the spatial distribution information of small and weak targets. Aiming at the problem that small target features are difficult to retain in the deep layer, a bottom-up multi-scale feature fusion module is designed to ensure that the low-level information of the small target can be retained in the high-level feature map with the attention mechanism between the feature maps at different levels. The test results on the public dataset NUAA-SIRST verify that the algorithm proposed in this paper has achieved better detection performance compared with the existing algorithms, while taking into account the optimization of accuracy and recall.