Search

Based on the YOLO v3 target detection framework, this paper trains and learns the public remote sensing image aircraft target data, and optimizes the defects of the YOLO v3 loss function, and introduces the IoU (intersection ratio) between the ground-true box and the prediction box, experimental results show that the precision, recall ratio and F1 value of the YOLO v3 model for aircraft target detection in remote sensing images are 95.12%, 86.21% and 0.9045, respectively. Compared with the previous ones, the network precision, recall rate and F1 value of the optimized loss function have been improved by 12.57%, 5.11% and 0.0863 respectively.

In this paper, an improved encoding scheme for LT codes is designed to reduce the error floor. The main idea of the proposed LT code is to classify the information nodes according to their degrees. Then the low-degree information nodes will be forced to frequently connect to check nodes, which enables all information nodes to obtain enough degrees to be restored successfully. Besides, the information degree distribution and convergence behavior of the proposed scheme are analyzed. Simulation results show that the proposed LT code can significantly reduce the error floor, and outperform the other two improved schemes, particularly for short length codes.