Research On Bolt Defect Detection Of Transmission Line Based On PCL Weakly Supervised Model

There are a large number of bolts in transmission line.It is of great significance to timely detect defective bolts for the safe operation of transmission line.Bolt defect detection based on fully supervised model needs object-level annotation,and the bolt image has the characteristics of small and many objects,object-level labeling will consume a lot of human,material and financial resources.In order to reduce the dependence on object-level annotation,this paper studies the transmission line bolt defect detection method based on proposal cluster learning(PCL)weakly supervised model,only uses image-level annotation to realize bolt defect detection,and innovatively improves the PCL model.Aiming at the problem of small target and difficult examples in bolt image,and the lack of bounding box regression mechanism in PCL model,a weakly supervised detection method of transmission line bolt defects based on improved PCL model is proposed in this paper.Firstly,the channel attention mechanism is introduced into the backbone to extract fine features of bolt objects.Secondly,the traditional classification loss function is weighted,and the balance coefficient of positive and negative example and the adjustment factor of hard and easy example are introduced into the traditional classification loss function to improve the learning degree of the model to the bolt object.Finally,the multi-task learning idea of full supervision is integrated to make the weakly supervised model have the ability of bounding box regression.The experimental results show that compared with the finetuning PCL model,the detection accuracy of the improved model is significantly improved on the datasets V1 and V2,which verifies the robustness of this method.On the basis of channel attention,spatial attention is added to enhance the regional connection of bolt object context and further help locate bolt object.According to the unbalanced detection performance of the PCL model on the dataset V1,the evaluation index of average detection precision difference among classes(ADPD)is defined,and a self-adaptive weighted loss function is proposed to dynamically adjust the learning degree of the model to different classes of examples,improve the detection accuracy of bolt defect and balance the detection accuracy of different classes,and the model is evaluated more comprehensively with the dual indexes of ADPD and mean average precision(m AP).Experiments show that the PCL model under the dual evaluation indexes shows a stronger ability to detect bolt defect.