A Trajectory Filtering Algorithm for Infrared Small Target Detection

The identification of faint and small targets in infrared imagery has consistently been a difficult task, mainly because of the poor signal-to-noise ratio and the scarcity of prominent texture features that can be associated with these targets. In this work, we introduce a Trajectory Filtering Algorithm aimed at enhancing the detection precision of faint and small infrared targets amidst challenging background conditions. The algorithm is integrated into a U-shaped deep learning network. By incorporating an external attention mechanism at each downsampling layer, global contextual information is effectively captured, thereby strengthening the extracted features. In the multi-frame dataset, a filter is added according to the trajectory to cover the non-target area of the picture, and the performance is better in the multi-frame data set. The experimental findings demonstrate that the proposed TFA framework achieves outstanding effectiveness in identifying small infrared targets within dynamic and cluttered backgrounds. Compared to other methods, TFA shows superior performance, as indicated by a more favorable Receiver Operating Characteristic (ROC) curve, a significantly lower false alarm rate, and an impressive Area Under the Curve (AUC) score of 0.9892.