Automatic horizon tracking method based on knowledge self-distillation

In recent years, deep learning has achieved a large number of successful cases in the seismic horizon tracking; however, most of the existing studies apply more complex network structures to improve the horizon tracking accuracy, which significantly increases the model complexity and training time. In order to improve the tracking accuracy without increasing the model complexity, this paper proposes a knowledge-based self-distillation method for horizon tracking. In the first step, the teacher model is obtained based on UNet training to obtain the prior weights; in the second step, the untrained UNet is used as the student network, and the teacher model is used to guide the student network for knowledge distillation training to improve the knowledge transfer performance. After synthesis and actual data testing, using the knowledge distillation method for horizon tracking improves the horizon tracking accuracy without increasing the model complexity, and provides a new method for automatic horizon tracking.