Artificial intelligence has learned to look for Chinese anti-aircraft systems

Center for geospatial intelligence University of Missouri using the methods of deep learning have developed an algorithm that is capable of satellite or aerial images to find Chinese anti-aircraft missiles. According to scientists, using their algorithm make it possible to handle reconnaissance survey in 80 times faster than humans. The work of researchers published in the SPIE Journal of Applied Remote Sensing, and a brief statement of the results of Aviation Week.

Currently, the processing of intelligence information is conducted by specially trained professionals who knew how to make a quick search of various important objects in photos and videos. To search for air defense systems, for example, are used, including typical signs, which is possible with high probability to speak about their location. For example, China is the place with the placement of such complexes in the pictures you can learn, for example, a typical circular arrangement of machines (and even an atypical location).

Center for geospatial intelligence — one of the American organizations responsible for the training of specialists in search of military equipment of the enemy on reconnaissance photographs. His experience in analyzing the specialists of the center and used for training the neural network. The researchers used a few for training convolutional neural networks: CaffeNet, GoogLeNet, ResNet-ResNet and 50-101. Training of neural networks was carried out on the well-known Chinese antiaircraft installations and photographs tipichnyh and atypical locations.

After training the neural network GoogLeNet showed the best average recognition result for images with the determined level of confidence in the end result more than 70 percent. At the same time ResNet-101 showed the best performance with high result with a confidence level less than 70 percent. Verification of the trained networks was done on unknown images. These same pictures were offered to specialists on the detection of anti-aircraft missile complexes. In the end, the neural network with an accuracy of 0.9 found anti-aircraft for 42 minutes. In humans, the figures were 0.9 and 60 hours.

Translated by «Yandex.Translator»