Temporal-Spatial Information Fusion Network for Multiframe Infrared Small Target Detection

The detection of small targets in infrared technology is of paramount importance in sectors like early warning and aerospace engineering. However, existing detection methods primarily focus on spatial information in single frames, neglecting the rich temporal information present in sequential frames. Our solution proposes a multiframe infrared small target detection method based on the fusion of temporal-spatial information, achieving the interaction of the reference frames and the current frame. First, to improve the capture features with sufficient distinguishability in the time domain, a temporal feature refinement transformer (TFRT) based on deformable attention is constructed, obtaining fine-grained features with a global temporal context representation. Furthermore, a spatial-temporal feature fusion transformer (SFFT) is designed based on cross-attention mechanism. This module adaptively fuses fine-grained temporal features from the reference frame with spatial features from the current frame, facilitating cross-frame spatial-temporal information interaction and enhancing the spatial features of small targets. Finally, a 3-D depthwise separable convolution (3DDSC) and 2-D depthwise separable convolution (DSC) is introduced to establish a weight-sharing temporal and spatial feature extraction network, respectively, thereby reducing the computational complexity and parameter count of the model. Experimental results on public datasets demonstrate that by effectively integrating temporal and spatial information, the model exhibits stronger anti-interference capabilities. It can effectively suppress false alarms while ensuring target detection, thereby reducing computational resource consumption.