Hyperspectral Image Target Recognition Based on YOLO Model

Jia-ju Ying,Shi-qu Yin,Wei-yong Yang,Hui Liu,Xu Li

DOI: https://doi.org/10.1117/12.3016421

2024-01-01

Abstract:Hyperspectral image target recognition, as an important component of computer vision, has broad application prospects in the military field. We conducted research on hyperspectral image target recognition technology based on deep learning, with typical military targets as the main recognition objects. A target recognition algorithm suitable for hyperspectral images is proposed based on the YOLO network model. To address the issue of the YOLO algorithm not being sensitive enough to small targets, the calculation formula for the loss function in the algorithm has been adjusted, and a multi-scale training method has been added. In response to the phenomenon of missed detections when the YOLO algorithm faces multi-objective clusters, five anchor boxes are added to each grid. In addition, the network convergence speed is accelerated by adding a BN layer; Introducing residual networks effectively prevents network degradation; Adding 1*1 convolutional layer can better extract image detail features and improve network recognition accuracy. The hyperspectral image target dataset is obtained by using hyperspectral imaging instruments to shoot multiple targets from multiple angles and directions under different lighting conditions, and labeling them. Divide the data into training and testing sets for training. The trained model was used for recognition experiments on vehicles in hyperspectral images. The experimental results showed that the model met the real-time requirements for target recognition in hyperspectral images, and performed well in the recognition performance of small targets at long distances.