Siamese YOLO V5 with Structure coefficient for object-level change detection

Abstract

Change detection is a task to identify the location and category of change objects in the reference image and the test image within a specific time interval, that can reduce workload and increase efficiency and reliability in applications such as foreign object intrusion, equipment status monitoring, building or natural resource monitoring, military anomaly monitoring, and so on. To deal with the complicated noise such as dithering, weather, lighting, shadows and background noise in the actual situations, this research combined the Siamese network in conjunction with advanced object detection for object-level change detection. The Siamese network with Structure coefficient is used to extract the fusion difference information between the reference image and the test image to be measured to resist the registration error and unrelated interference between images, which is used for YOLO V5 to detect the effective rectangular boxes and category of the change objects. Four public datasets of different scenes include LEVIR-CD, VL-CMU-CD, AICD-2012 and CDNET-2014 are used in multiple comparative experiments, and the experimental results proved that our method achieved higher accuracy than existing object-level methods and less false detections than existing pixel-level methods.