基于仿生视觉骨干的级联蒸馏输电线路目标检测模型

Target Detection Model of Cascade Distillation Transmission Line Based on Bionic Visual Backbone

针对输电线路无人机智能巡检中多个小目标缺陷检测精度低且复杂环境下缺乏鲁棒性等问题,提出基于仿生视觉骨干的级联蒸馏输电线路目标检测模型。首先,设计2个共用骨干网络的高效模块构建级联蒸馏结构:动态锚框蒸馏模块和动态锚框提纯模块,前者负责粗略寻找可能存在目标的区域,后者负责细化该区域,从而提出一种新的渐进式目标检测方法,解决输电线路中受遮挡目标的漏检问题。其次,构建仿生特征提取骨干网络,通过模仿生物感受野充分聚合上下文信息,提升骨干网络对线路中尺度较小目标的特征提取能力。然后,设计周边视觉模块,利用人类独有的视觉机制控制局部信息的交互强度,克服输电线路中复杂多变的背景对模型的干扰。最后,构建生成了输电线路目标的数据集,并通过仿真实验验证了模型的泛化性和鲁棒性。

Aiming at the problems of low detection accuracy of multiple small target defects and lack of robustness in complex environments in intelligent inspection of transmission line UAVs,a target detection model of cascade distillation transmission line based on bionic vision backbone is proposed.Firstly,two efficient modules sharing the backbone net-work are designed to construct a cascade distillation structure,namely,dynamic anchor frame distillation module and dynamic anchor frame extraction module,the former is responsible for roughly finding the area of possible targets,and the latter is responsible for refining the area,so as to propose a new progressive target detection method to solve the problem of missing detection of occluded targets in transmission lines.Secondly,a biomimetic feature extraction back-bone network is constructed,and the feature extraction ability of the backbone network for smaller targets in the line is improved by fully aggregating the context information by mimicking the biological receptive field.Then,the peripheral vision module is designed,and the interaction intensity of local information is controlled by using the unique visual mechanism of humans to overcome the interference of complex and changeable backgrounds in transmission lines on the model.Finally,a dataset of transmission line targets is constructed,and the generalization and robustness of the model are verified by simulation experiments.