Research On The Image Instance Segmentation Algorithm Of Electrical Equipment Image Based On Deep Learning

The safe operation of substation is the key factor to ensure the stability of the power system.Regular inspection of power equipment and power lines can make potential safety hazards spotted in time and personal and property losses avoided caused by major power failures.The traditional manual patrol inspection has disadvantages of large workload,low efficiency and advert influences causing by the geographical environment.The unmanned aerial vehicle inspection system based on artificial intelligence technology has the characteristics of close observation,safety and efficiency,which can overcome the shortcomings of manual patrol inspection such as low automation and lack of data accumulation and communication.The key technology of intelligent inspection system of electrical equipment is automatic identification and fault diagnosis of inspection images,and the key of these two tasks is the accurate segmentation technology of equipment targets in the images.This research proposes a method of identification and instance segmentation of substation electrical equipment images based on improved Mask R-CNN algorithm.Firstly,aiming at improving the poor effect of image edge information and low accuracy of small target recognition in Mask R-CNN,a diversified feature pyramid network(DFPN)is proposed.Secondly,a diversified atrous feature pyramid network(DAFPN+)is proposed by introducing the atrous spatial pyramid pooling module into the DFPN,which can overcome missing detection by the scale change.Furthermore,Then,1730 images are labeled and an outdoor electrical equipment segmentation dataset(OEE)is established,based on 3996 images of six types of outdoor electrical equipment collected during 5-year manual inspection of urban substations.Finally,by testing this network and traditional Mask R-CNN on OEE,the results show that our network can accurately identify six types of typical electrical equipment including arrester,current transformer,breaker,disconnector,reactor and insulator.The recognition accuracy is improved by 7%,and the segmentation accuracy is improved by 8%,so that the small-scale targets in images can be effectively identified,and the missed detection rate is reduced.