Research And Application Of Image Recognition Algorithm For Power Equipment Based On Deep Learning

With the improvement of automation,inspection robots are more and more widely applied into substations,facilitating the acquisition of quantities of power equipment images.However,the low accuracy in recognizing the reading of pointer-type meters and transformer small components in these substation inspection images has become a bottleneck.In recent years,deep learning algorithms have made important breakthroughs in the field of image recognition by taking advantage of big data.Therefore,this thesis analyzes the characteristics of power equipment images and studies the automatic recognition of the pointer meter readings and transformer small components.The achieved results are as follows.(1)An automatic recognition method of pointer meter reading based on Faster R-CNN object detection and U-Net image segmentation techniques is proposed.Firstly,Faster R-CNN is introduced to detect dial and pointer area in the images as well as to classify meters.Then with the concept of replacing graphic detection algorithm with image segmentation technology,the structure of U-Net is improved according to characteristics of the meter images so that scale lines and pointers can be effectively extracted.U-Net dice loss function is constructed for the problem of segmentation category imbalance.To correct the rotation of meters,a perspective transformation method based on segmentation information is proposed.The proposed method can greatly improve the recognition accuracy of pointer meter reading under complex background.(2)As the above recognition method of pointer meter reading has too many intermediate links and its speed is quite slow,a new pointer meter reading recognition method based on the fusion of attention mechanism and convolutional neural network is proposed.After integrating convolutional block attention module and U-Net into the convolutional neural network,the model can be guided to focus on features that are more relevant to reading recognition of pointer-type meters.The soft stage-wise regression method is used to better solve the boundary problem caused by the pointer stopping between scale lines.The designed model is simple in structure and has an advantage in recognition speed while effectively avoiding the accumulation of errors caused by intermediate links?(3)To improve the recognition accuracy of transformer small components,a recognition method based on Retina Net is proposed.First,the object detection network Retina Net is improved,and the fusion feature map with higher resolution is added to solve the problem that transformer small components contain too little pixel information.Then,a probability correction method for transformer small components based on location correlation is proposed,which can modify the probability of small component detection boxes according to the location correlation between large components and corresponding small components,to avoid the interference of other similar components to the target component recognition.Finally,the proposed recognition method is verified by the actual substation inspection images,and the result shows that the method has significant advantages both in the recognition accuracy of three types of transformer small components and the overall recognition accuracy.