DA-YOLOv7: Velocity Spectrum Pickup Based on Domain Adaptation

The automatic picking of velocity spectrum based on object detection has proven its effectiveness, but these models are limited to the current training work area and cannot get effective inference in other work areas. To improve model generalization and achieve the automatic picking of velocity spectrum in other work areas without enough velocity spectrums and labels, our paper designs a Domain Adaptation velocity spectrum picking model DA-YOLOv7, and optimizes the network to improve the picking accuracy. At the same time, in order to address the domain gap, training pseudo images are generated through image-to-image translation and used as inputs for model training. Finally, post process the results through the constrained KNN anomaly detection method, screen and eliminate the outlier, so that the results are closer to the manual picking up. By comparing with the SOTA domain adaptation networks, it is proved that DA- YOLOv7 is superior to the existing model in precision and visualization.