Improved YOLOv5 with Transformer for Large Scene Military Vehicle Detection on SAR Image
Yi Sun, Wenna Wang, Qianyu Zhang, Han Ni, Xiuwei Zhang
Abstract
With the development of SAR technology, large scene object detection on SAR images has attracted more and more attention. Exiting large scene object detection is mainly based on the CNN network, which limits the obtaining of global context information. On the other hand, due to the high acquisition cost of SAR images, there are no existing public datasets in military vehicle detection. To solve these problems, we adopt the Transformer module to construct the neck block based on YOLOv5. This design can gain global context information, and also has better performance for small objects detection. Furthermore, to achieve the detection of large-scale military ground vehicles, we construct a dataset based on the MSTAR dataset, named LSGVOD. Extensive experiments have been conducted on LSGVOD, and experimental results show that the proposed method greatly improves detection accuracy. Compared to other methods, it achieves the best accuracy with 93.3% mAP.