Infrared Small Object Detection Based on Spatial–Frequency Enhancement and Manifold-Constrained Multiresidual Redistribution

Infrared small object detection (ISOD) plays an important role in remote sensing and security applications, yet it still faces challenges. Infrared small targets exhibit weak feature responses in the spatial domain, which can be compensated by incorporating frequency-domain information. In addition, the identity mapping in residual connections of existing networks causes the continuous propagation of background features, thereby weakening the representation capability of small targets. To address these issues, we propose a frequency-guided multiresidual manifold redistribution network (FMRNet). Specifically, to tackle the weak response of small targets in the spatial domain, we propose the spatial–frequency modulation (SFM) module. This module uses a lightweight multilayer perceptron (MLP) to adaptively reweight amplitude, highlighting sparse high-frequency target energy, and fuses it with spatial features to improve target separability. To address the excessive propagation of background features in residual connections, we propose the manifold-constrained multiresidual redistribution (MMR) module as the core feature extraction component of FMRNet. This module constructs multilevel, multiscale feature streams through multiple bottleneck blocks and maps residual connection space onto the manifold of doubly stochastic matrices to dynamically redistribute multiresidual streams, thereby effectively suppressing the propagation of redundant background features. To address information loss caused by downsampling, we propose the high-fidelity parallel downsampling (HPD) module. This module enhances semantic and spatial features via convolutional and pooling branches, with concatenated outputs expanding channels and retaining more target information. Experimental results show that our method outperforms state-of-the-art (SOTA) methods on public datasets.