SceneNet: A Multi-Feature Joint Embedding Network With Complexity Assessment for Power Line Scene Classification

Power line extraction is not only crucial for UAVs obstacle avoidance, but also a fundamental step for fault diagnosis of power lines. Therefore, achieving robust and accurate extraction of power lines in aerial images is essential to enable intelligent UAVs inspection. Unfortunately, power line extraction is an extremely challenging task, and all the current methods attempt to utilize a single model to solve the problem of power line extraction in complex and variable scenes. This results in insufficient generalization ability and suboptimal computational efficiency. In this work, we propose a power line scene classification network based on complexity assessment, named SceneNet, which can provide a solution for tackling power line extraction challenges. Firstly, we propose a human-machine hybrid reasoning model to obtain the ground truth of image complexity reasonably and build the first benchmark dataset that can be used for automatic classification research of power line scenes. Secondly, we propose an improved StyleGAN3 model and loop transfer learning strategy for data augmentation. Most importantly, the SceneNet comprises a multi-feature joint embedding module and a feature encoding-decoding module. On the one hand, it achieves the multi-level fusion of artificial features and high-dimensional semantic features. On the other hand, we use a self-attention mechanism to enable full use of the contextual association between each block of the fusion feature map. The SceneNet has successfully achieved the mapping and pattern recognition between the abstract concept and the concrete features. Experimental results demonstrate that the SceneNet is obviously superior to the existing 12 state-of-the-art models, and it provides guidance and delineation of applicable scenes for power line extraction methods.