Detection of space infrared weak target based on YOLOv4

In the problem of space infrared weak target detection, traditional algorithms such as correlation template method and center of gravity, etc. have a low discrimination rate and high data quality requirements. To solve this problem, the space infrared weak target detection algorithm based on improved YOLOv4 is proposed in this paper. The algorithm first establishes corresponding data sets for different infrared targets in space. Based on YOLOv4, a special neural network framework for space target detection tasks is established. The k-means clustering algorithm is used to reconstruct the prior frame, and multi-scale fusion is designed according to the characteristics of infrared weak targets to improve the detection accuracy of weak targets. Finally, COCO data set and the infrared image data set collected in the laboratory are used to train the algorithm and test. The test results show that the improved algorithm has a significant improvement in the accuracy of detection compared with the YOLOv4 algorithm. Its average accuracy (AP) can reach more than 93.25%, and the detection speed has reached 38.99 ms/frame, which verifies that the effectiveness of target detection satisfies the needs of space infrared weak target detection tasks.