Transmission Line Abnormal Object Detection Method Based On Two-stage Deep Network

The abnormal object detection of transmission lines plays a very important role in improving the safety,reliability and stability of transmission system.Transmission lines mostly exist in the field,the environment is complicated,and manual inspection is time-consuming,labor-intensive,and poor in effect,which cannot meet the needs of abnormal object detection.The object detection algorithm based on deep learning has high detection accuracy,fast speed,and good tolerance for complex scenes,However,it has not effectively designed the problems of large changes in scale of abnormal objects of transmission line,many small objects,low light,partial occlusion,etc.resulting in slow recognition speed,susceptibility to environmental interference,frequent false alarms,and frequent false alarms.In response to above problems,this paper proposes a two-stage detection method for abnormal objects on transmission lines.The main research contents are as follows:A method for detecting abnormal objects in transmission lines generated by multi-scale feature enhancement and feature-guided is proposed.Multi-scale features are extracted by FPN,and prediction is made on feature maps of different scales,which solves the problem of many small objects in transmission line dataset.Global network is used to enhance multi-scale features extracted by FPN,filter noise in feature map extracted by FPN,and extract clearer and more representative object features,which can adapt to fuzzy and unclear characteristics of transmission line dataset.The feature-guided candidate box generation network generates a tighter bounding box,which solves the problem of a tight bounding box of individual categories(such as tower cranes)in transmission line dataset.At the same time,a multi-task loss function is used to improve prediction accuracy and generalization ability of network,and improve performance of abnormal object detection.The AP of this method on MS COCO dataset reached 41.7%,and at the same time,the detection accuracy on transmission line dataset reached 77%,and good detection results were obtained.A global feature enhancement and adaptive regression method for abnormal object detection in transmission lines is proposed.In the feature extraction stage,FPN and HRNet are combined to extract multi-scale high-resolution features.The global network enhances features to obtain higher-quality features,reduces the impact of poor transmission line dataset image quality on feature extraction,removes noise that is not related to object,and effectively separates background from foreground.In the training phase,positive and negative sample equalization is introduced to mine difficult samples to improve detection performance,and adaptive bounding box regression loss is introduced to reduce influence of occlusion and boundary blur in transmission line dataset on improving network performance.AP on MS COCO dataset reached 45.3%,and detection result on transmission line dataset reached 79.6%,which is 2.6% higher than transmission line abnormal object detection method generated by multi-scale feature enhancement and feature guidance candidate frame.At the same time,the single image detection speed of this method reaches 0.21 s,which is nearly four times faster.