Oriented Target Detection Algorithm Based on Transformer
Zhizhong Xi, Jingen Wang, Yanqing Kang · 2021 4th International Conference on Artificial Intelligence and Pattern Recognition · 2021
Oriented target detection is an important task in the field of remote sensing target detection, which has great application prospects in geography, agriculture and military. However, the current popular algorithms of oriented target detection do not achieve the best in speed, accuracy and computational complexity. In this paper, an oriented target detection algorithm based on Transformer is proposed, which uses the backbone(CSPNet) and neck(FPN + PAN) of YOLOV5 model to realize Feature extraction and multi-scale feature fusion. The multi-scale feature map is input into Transformer module for classification and regression. By introducing the transformer module into the oriented target detection task, the attention mechanism modeling at the scene level can be realized, and the detection accuracy can be improved; Using multi-scale feature map for self-attention modeling can improve the detection performance of the model for targets of different sizes; In addition, the multi-layer feature map is cut by region in the Sequence encoding, which can reduce the calculation cost and improve the detection speed.