Infrared Small Moving Target Detection Based on Combination of 3D-Conv and 2D-Conv

Infrared dim and small targets usually have the characteristics of small target size, few pixels, and low SNR, which are easily submerged by background and noise. By analyzing the characteristics of infrared small targets, we notice that the motion characteristics of infrared targets are different from noise. In addition, 3D Conv in the spatiotemporal domain can effectively extract the motion features of infrared small targets. Based on the characteristics of infrared small target images, this paper proposes an infrared small target detection algorithm that combines 3D Conv and YOLOv5. In experiments, we achieved significant improvements in false alarm rate and detection rate when the number of parameters only increased by 0.3% compared to yolov5s neural network model. In the presence of noise interference, our proposed model can reduce the number of false alarms by at least 40% and achieve a precision rate of over 96%, which proves that the model has good robustness. The experiment proved that our proposed algorithm combining 3D conv and 2D conv can effectively detect Infrared dim small targets, with good detection rate and low false alarm rate.