Infrared target detection using deep learning algorithms

Automatic target recognition is critical in infrared imaging guidance. However, since the diversity of the environment, the infrared data are often complex and difficult to analyze accurately. We proposed a deep learning infrared target detection framework based on transposed convolution and fusion modules (TF-SSD). Compared with one-stage detector YOLOv5 and two-stage detector Faster R-CNN, TF-SSD has three highlights: (1) using the visualization method to revise the network structure and improve the training efficiency; (2) using transposed convolution operation to increase feature extraction ability and detection efficiency; (3) using multi-scale feature fusion models to realize the skip connection between the high-level network and the low-level network. Experimental results on our dataset of six common flight attitudes demonstrate that the maximum average precision (AP) is 90.9%, the minimum average precision is 79.8%, and the overall mean average precision (mAP) is 85.7%. It is confirmed that our proposed TF-SSD system can effectively recognize infrared targets.