Scientists at the Ural Federal University have developed a neural network that detects cracks in bridges, roads, and buildings in seconds. The technology speeds up inspections by 100 times and reduces the risk of accidents. The model's accuracy is 88.7%, and it has been tested on images from Russia and China.
Manual inspections take 1–2 hours, while the algorithm analyzes images in seconds. The system processes up to 232 frames per second and has only 2.51 million parameters, allowing it to be used in drones. To improve accuracy, scientists added the SimAM attention mechanism and the Concat_BiFPN module, which reduce false positives on shadows or dirt.
The neural network is now being adapted to work with drones on the Jetson platform and infrared cameras. This will allow for the detection of hidden defects at night or under contamination. To save computing resources, the C3Ghost module was applied, reducing the model's parameters by 16% without loss of accuracy.
According to Rosstat, 60% of structural failures occur due to the untimely detection of cracks. Traditional diagnostic methods require the participation of up to three specialists per object. The first crack monitoring system using neural networks appeared in Japan in 2018 but depended on powerful servers. The Ural Federal University's development is 10 times lighter than foreign AI-based analogs, which often do not support real-time operation.
Read more on this topic:
Ryazan Radio Engineering University Develops New Optical Communication Modules for UAVs
Delivery Robots Will Appear on the Streets of 37 Regions of Russia