Multiobjective Neural Architecture Search for Power Line Recognition

Power line recognition is critical for the safety of the power grid. While the traditional recognition methods rely on inspection workers. In recent years, several studies have tried to use convolutional neural networks to replace the burden of labor. However, all of them only focus on the recognition accuracy while ignoring the model size, which make deploying drones impractical. To address this, we propose a multiobjective neural architecture search method for power line recognition, termed PLR-MNAS. PLR-MNAS can automatically design high performance and lightweight CNN architectures for power line recognition using the multi-objective search strategy. In addition, zero-cost proxies are performed on the evaluation process of PLR-MNAS, significantly reducing the search cost. Experiments with the with the PLR-PLD dataset show that PLR-MNAS can perform better than peer competitors. Specifically, PLR-MNAS achieves 100% accuracy with only 0.099 M parameters.