Artificial Intelligence has learned to look for Chinese anti-aircraft complexes.

The Center for Geospatial Intelligence of the Missouri University, using deep training methods, developed an algorithm capable of finding Chinese anti-aircraft missile systems on satellite or aerial photographs. According to scientists, the use of their algorithm will allow to process reconnaissance survey 80 times faster than people. The work of the researchers is published in the "SPIE Journal of Applied Remote Sensing", and a brief summary of it is given by Aviation Week.

Source    Typical locations of anti-aircraft complexes.

Currently, intelligence data processing is conducted by specially trained specialists who know how to quickly search for various important objects in photographs and video recordings. For the search for anti-aircraft complexes, for example, typical signs are used, for which it is possible to speak with high probability about the locations of their location. So, in China, places with the placement of such complexes in the photographs can be found, for example, on the typical circular arrangement of machines (but also atypical location).

The Center for Geospatial Intelligence is one of the American organizations responsible for training specialists in the search for enemy military equipment on reconnaissance photographs. His experts in the analysis of photos specialists of the center used to teach the neural network. The researchers used several convolutional neural networks for training: CaffeNet, GoogLeNet, ResNet-50 and ResNet-101. The training of neural networks was carried out on photographs of known Chinese antiaircraft installations and photographs of typical and atypical locations.

Source   Typical location of anti-aircraft complexes. A - places for launchers, B - mobile launchers, C - ramps, D - ramps for radar installations, E - places for launchers, F - launchers, G - ramps, H - concrete screens for flame protection from missile engines.

After training, the GoogLeNet neural network showed the best average recognition result for images with an established level of confidence in the final result of more than 70 percent. At the same time, ResNet-101 demonstrated the best performance with a high result with a confidence level of less than 70 percent. Checking the trained networks was carried out on the images they did not know. The same photographs were offered to specialists in the detection of anti-aircraft missile systems. As a result, neural networks with an accuracy of 0.9 found anti-aircraft installations in 42 minutes. In humans, these rates were 0.9 and 60 hours.

In mid-July this year, researchers from the University of Granada in Spain, using computer-assisted training methods, developed software that can detect a gun in real-time in a video recording or in a video broadcast. The new software will make it possible to realize the detection of small arms only by means of video surveillance. The weak side of the program, according to the developers, is only its inability to detect hidden under the clothes of weapons.

When creating the program, scientists used a pre-trained neural network. Her training in object recognition was carried out on the basis of images ImageNet, which includes about 1.3 million photographs of objects of about a thousand different classes. Exact training in the recognition of weapons was carried out on three thousand photographs of weapons prepared by researchers. As a result, the researchers received a program that can accurately determine the weapon on video recordings with quite high accuracy. Accuracy of the algorithm work compositionkztn 96.6 percent.

Source