Automated Real-time Processing of Satellite Images Without Human Intervention Using Explainable AI Models

The Space Vehicles Directorate of the US Air Force and Space Force is interested in image processing on-board satellites for real-time detection of objects of interest. Obviously, such analysis, being on a satellite, must be performed without any human assistance and in a reliable way. Teuvonet has the technology that can provide that capability. That technology is explained using the overall AI technology framework that DARPA has provided over the years through several of its programs. Deep learning models, that are used for image recognition nowadays, are complex, hard to understand and therefore too risky to deploy in defense systems. DARPA launched an Explainable AI (XAI) program in 2016 asking for the development of more transparent and explainable models that can be deployed with confidence in defense systems. For computer vision, DARPA wanted object prediction to be based on the verification of parts of the object. For example, to predict that there is a cat in the image, it wanted the system to verify that it found its claws, whiskers, fur and other unique parts of a cat. So, part-verification based object prediction is DARPAÆs conceptualization of Explainable AI (XAI) for computer vision. Arizona State University (ASU) has developed this type of XAI technology and has filed an international patent application (a PCT) application for it. Teuvonet Technologies, a faculty startup, has licensed this technology from ASU and is currently using it to create a variety of real-time image processing systems based on such transparent models. DARPA is also interested in having AI models that are robust against adversarial attacks and started the GARD (Guaranteeing AI Robustness Against Deception) program. We have found that these parts-verification based XAI models are naturally robust against adversarial attacks. DARPA also started the ANSR (Assured Neuro Symbolic Learning and Reasoning) program. In the ANSR Program, DARPA wants ōnew, hybrid (neuro-symbolic) AI algorithms that deeply integrate symbolic reasoning with data-driven learning to create robust, assured, and therefore trustworthy systems.ö TeuvonetÆs technology supports creating systems that the ANSR program needs. Image processing, in general, whether from satellites, drones, or CCTVs, requires humans to examine them and is, thus, very labor intensive. Using transparent and trusted Explainable AI models, Teuvonet is developing a technology for image processing (1) that would not require humans to continuously observe and verify the content of the images, and (2) where processing and analysis of the images, for object detection, can be done with high-speed in real-time on hardware housed in satellites, drones and CCTVs. Such a capability would also obviate the need to transmit large volumes of unimportant images to an external storage facility such as the cloud. In other words, in this mode of processing, only important images will be transmitted to and stored in an external facility.