Transmission Line Abnormal Object Detection Method Based On Deep Network

In order to ensure the reliability,stability and effectiveness of the transmission system,early detection and effective response to the possible risks of transmission lines are crucial.Deep learning-based detection methods can detect transmission line anomaly targets more efficiently,but there are problems such as slow detection speed,low detection accuracy,high false detection and leakage rate,and difficult deployment of edge-end devices due to the problems of large target scale changes,many small targets,complex backgrounds,uneven data set categories,and oversized detection models in transmission lines.In this paper,we propose a deep network-based transmission line anomaly target detection method to solve the above problems,and its main research includes:An abnormal target detection method for transmission lines based on feature perception enhancement and bidirectional refinement is proposed.An enhanced feature extraction network is designed.Enhancing feature perception is accomplished through the utilization of the module,thereby diminishing information loss in the feature extraction process and augmenting the retention of feature data of minor targets.To accommodate the target’s multi-scale alterations and minimize the interference of intricate background data,channel optimization and spatial optimization modules are incorporated to form a two-way feature fusion network.The network can obtain salient feature information with strong semantics and accurate location.The introduction of the balanced sampling technique into the region proposal network is intended to augment the sampling likelihood of challenging samples.The class imbalance issue can be alleviated by introducing the adaptive class suppression loss function as a classification loss.The method achieves an accuracy of 89.3 % on the transmission line dataset and can cope with various difficult situations.An abnormal target detection method for transmission lines based on cross-order residual enhancement and attention feature fusion is proposed.The feature extraction network of cross-order residual enhancement is constructed.The deep separable convolution is introduced and its convolution kernel is expanded.Ensuring the model’s lightness,the receptive field is broadened and the abundant feature data is taken in,thus augmenting the detection velocity and precision.A technique of feature fusion,based on an attention mechanism,is suggested for cross-scale data fusion,thereby resolving the issue of multi-scale disparity of minor objectives and augmenting the model’s detection proficiency.A detection head decoupled,coupled with a dynamic label allocation method with double weights,is finally implemented to enhance the training samples’ quality and the model’s detection performance.The transmission line data set’s experiments have demonstrated that the technique attains 90.5 % detection accuracy while lessening the amount of parameters.An abnormal target detection method for transmission lines suitable for edge-end equipment is proposed.The backbone network and detection neck are constructed by using the idea of reparameterization,which improves the reasoning speed without affecting the detection accuracy.Optimizing the SPPF module through CSP,the detection accuracy is enhanced while the number of parameters is minimized.A multi-layer feature fusion module is incorporated into the feature fusion stage to maximize the accuracy of small targets’ detection by utilizing the position information of the features.Additionally,a detection head without anchors is utilized and an anchor frame is included to facilitate training and further augment detection accuracy.No inference delay will be caused by the exclusion of the auxiliary module in the inference stage.Experimentation on the transmission line data set has yielded a detection accuracy of 83.2%,a detection speed of 609 FPS,and model parameters of only 4.7 M,which can fulfill the deployment needs of edge devices.