Research On Image Change Detection Technology In Substation Scene

Image change detection is a comparative analysis of the input reference image and test image,ignoring some minor or meaningless change areas,and finding the main change areas in these two images to give alarm information.It is an important application of image monitoring.The paper studies the image change detection technology in the substation scene.The substation inspection robot collects equipment patrol images on a preset track,compares the images of the same scene taken at different times,and detects changes in the state of power equipment,such as the opening and closing status of the electric box door,the opening and closing status of the knife switch,the damage of the instrument,or the intrusion of personnel,etc.The paper first analyzes the overall system of change detection.Regarding the problem of the offset between the image input to the system,For the registration problem between input images,the image registration module in the change detection system is introduced,which compares the image registration algorithm based on traditional SIFT and the image registration algorithm based on deep learning D2-Net,combined with experiments to analyze the performance of the performance of these registration algorithms.Experiments show that the image registration based on D2-Net performs better in scenes with strong illumination changes and weak textures.For the change detection module after registration,the paper introduces the existing change detection network based on deep learning,and analyzes the specific reasons for the low detection accuracy of the existing change detection network based on experiments.The paper designs a change detection network CD-Unet.On the basis of the combination of the Siamese network and the semantic segmentation network,the Siamese network feature extraction module based on ResNet-34 is designed,and the different fusion methods of the feature maps obtained from the Siamese network are analyzed and tested.In the output stage of the change detection,multi-scale feature fusion is performed to make the final output result more accurate.In the stage of training the change detection network,the paper proposes a training method using joint loss.Based on the above design,the CD-Unet’s F1 index on the CD-2014 dataset reached 0.9554,and the F1 index on the change detection dataset of the substation scene reached 0.9670,the average calculation time is only 14.49msThe paper also studied the algorithm of the process from the change detection result to obtaining the ROI rectangle frame of the change area.Under the premise of obtaining the ROI rectangle frame of the change area,the feature extraction module based on ROI Align was designed.The paper also proposed a changed category recognition network that shares some parameters with the change detection network feature extraction and finally achieved 99.68%recognition accuracy on the changed category recognition dataset in the substation scenario.the task of detecting the change area and recognize the change category can be completed in 14.68ms at the same time.