A local-patch and global attention network with multiscale asymmetry fusion for infrared small-target detection

Abstract With the development of image recognition technology, small-target detection has attracted wide interest in early warning systems due to the fact that it operates in real-time and has a high accuracy rate. To further improve the accuracy and robustness of small-target recognition in the field of infrared surveillance, based on the mechanism of global feature capture and local detail enhancement, we propose a local-patch and global attention network (LPGANet) with multiscale asymmetric fusion based on binary segmentation. This allows us to efficiently capture local-patch feature details and global position information of small targets for segmentation between targets and backgrounds. In particular, the multiscale asymmetry fusion module utilizes a combination of deep and shallow semantic features to achieve pixel preservation at key up-sampled locations. By fusing multiscale feature maps at different levels, the target detection accuracy can be effectively improved. The microscopic recognition module (MRM) can effectively perform feature extraction for target weak edges, and by introducing MRM into the global network, which can optimize the extraction ability of dim and dark target edges, and ensure the robustness of network detection. Extensive experimental results show that LPGANet is effective in recognizing small infrared targets from the background and that it demonstrates high accuracy and stable detection results.