Target Neighborhood-Aware Enhanced Small Infrared Target Detection Algorithm*

The low detection rate of small targets is a common problem in CNN-based general object detection models. In particular, the detection of small targets in infrared imaging is more challenging due to the high false detection rate caused by low contrast in complex backgrounds. Insufficient perception of contextual information around small targets in convolutional operations is the main reason for false detection. To address this issue, this paper proposes a transformer-based method for enhancing neighborhood perception in infrared small target detection. For two-stage object detection methods, this paper first proposes candidate region resampling to collect as much contextual information as possible for the target regions generated by RPN. Then, a transformer encoding structure is utilized to enhance the global perception ability of features for the neighbor-hood. Finally, a classifier is designed to evaluate the differences between potential target regions and their neighborhoods, which is used to determine whether the RPN candidate region contains a real target. With detection accuracy, recall rate, and F1 score as evaluation metrics, the proposed detection framework achieves 92.39%, 97.28%, and 93.63% on the MDvsFA dataset and 95.25%, 98.35%, and 96.13% on the ALCNet dataset. The proposed infrared small target detection algorithm, which enhances neighborhood perception, combines the advantages of traditional algorithms and transformer structures to achieve more accurate results.