Research On Object Detection Of Electric Power Fittings Images Based On Deep Learning

Electric power fittings are important metal parts that connect and combine power system devices and require regular inspections.Traditional manual inspection is not suitable for the development trend of intelligent grids due to its low efficiency and high cost.Intelligent inspection technology has become prevalent in continuously constructing innovative power systems,but much inspection image data still needs further analysis.This study proposes a UNet-DB＿ECA method base on deep learning to realize pixel-level electric power fittings image detection,which can solve the above problems and facilitate deployment in edge devices.The study expects to provide a new way of thinking about applying deep learning in this field.The main research work of this study is as follows:(1)To realize the lightweight,intelligent detection of electric power fittings images,this study proposes a UNet-DB＿ECA method base on the lightweight improvement of the U-Net network.Firstly,Decreasing the U-Net network channels shrinks the model size and speeds up the training.Secondly,the network embeds a channel attention module before upsampling to enhance channel utilization.Finally,The network performs batch normalization operations,and part of the Re LU activation function transforms into a Hard-Swish activation function,improving the neural network’s performance.The experimental results show that the UNet-DB＿ECA network can better detect the image of electric power fittings.(2)Since the activation function has inconsistent effects on different types of electric power fittings image datasets,this study researches the influence of the activation function on the neural network,analyzes three different placement methods of the activation function through experiments,and then explores ten different methods.The influence of type activation function on the UNet-DB＿ECA network provides a basis for selecting activation function in subsequent deep learning detection methods.(3)To apply relevant theories in practice,this study designs an intelligent humancomputer interaction system for image object detection of electric power fittings.The system provides data analysis and management functions through two graphical user interfaces of single-image detection and batch detection and realizes the visualization and efficiency of detection tasks,which is more suitable for actual engineering scenarios.In summary,this study proposes the UNet-DB＿ECA method for the lightweight,intelligent detection of electric power fittings images,and experiments verify the method’s feasibility.