An Improved Small Infrared Target Detection Algorithm Based on Yolov5

Aiming at accurately, robustly detecting the small infrared targets, an improvement to the current deep-learning detection method YOLOv5 is needed. Firstly, feature fusion and attention module are introduced to better fuse feature maps across layers and augment the features using the attention mechanism. Secondly, by adding Bi-FPN in the neck to integrate semantic information from high to low layers to better detect small targets. Finally, to avoid the cases that GIOU could not recognize, the CIOU is used to better train the network. Experiments are performed on the IRSTD-1K small infrared targets dataset. The improved method gets higher 0.001 in Precision, 3.6% increase in mAP50, 2.15% increase in mAP50-90, compared with YOLOv5.