Abstract:
In order to realize safe and efficient ice-cover detection, machine learning models based on computer vision have emerged. However, with the continuous expansion and development of the power grid scale, the traditional ice-cover detection models face the challenges of the difference in data distribution and the dramatic increase in data volume, which seriously affects the detection efficiency. To this end, we introduce a continuous learning strategy and proposes a layer-aware prompting for visual continual learning(LP-VCL). The method effectively overcomes the problem of data distribution differences and copes with the dramatic increase in data volume by learning soft cues for each layer of encoder, thus improving the robustness and generalization ability of the model. To validate the effectiveness of the proposed model, we compare it with existing state-of-the-art methods and test it under different continuous learning settings, including class increment, domain increment, and task-independent settings. The experimental results show that LP-VCL outperforms the baseline method in two key metrics, namely average accuracy and forgetting rate, especially in the class-increment and domain-increment settings, which proves the superiority and practical value of the method.