Artificial intelligence is an emerging technology with a large potential in military applications. Dr. Dijk’s goal is to improve the automatization of different military tasks, especially detection, recognition, and identification (DRI) of targets and threats using optronics sensor systems as part of the total sensor suite, on new and current military platforms. In this talk, she will provide an overview of the added value and challenges of applying AI for EO sensors in military applications for data-driven approaches such as deep learning.
Deep learning (DL) is a fast-emerging technique for automating a large variety of tasks with promising results. In the field of optronics, so-called convolutional neural networks are used to automatically extract information from images and videos for classification and object detection. Many models are trained on publicly available datasets and ready to be used by anybody. These datasets consist of images and their corresponding label: the class or the location of the object in the image. The advantage of these models is that the important features to distinguish different classes or to find the relevant objects are already learned. However, this only works for imagery similar to that of the data used to train this model. For typical defense applications the systems the models need to be updated with specific military imagery, such as infrared or image intensifier images, or for typical targets in their environments. We present examples where DL technology can be applied to improve situational awareness using imagery from different platforms: from an airplane and static from land, and in different environments, both land and sea.
Dr. Dijk identifies five challenges: 1) application in the military domain is different than application in the civil domain, 2) there is (too) little training data available, 3) there is a need for trust in the results of the system, 4) there is a need for online, adaptive systems and 5) there is a need for edge processing systems. To tackle the first two challenges and improve the models, recently developed approaches such as transfer learning, learning with less labels, learning by detecting parts of the object, data simulation and data augmentation and zero-shot learning can be applied. Trust in the system can be achieved using different approaches, e.g., by explaining the results to the operator or by adding knowledge, e.g., from the operator to the system. The last two challenges need different hardware and training solutions. Awareness of the challenges of AI and possible solutions will improve the applicability of AI developments in military applications.
Judith Dijk obtained a Ph.D. in applied physics at Delft University of Technology. Since 2003 Judith works at TNO, a Dutch research institute. Her research focuses on research and development of imaging systems for defense applications, both on the hardware and on the software needed for these. In recent years, this includes insight in how artificial intelligence can be best used within different defense applications, and what is needed in the technology development to achieve these capabilities. As one of the lead scientists of TNO’s AI program, Judith is at the center of low TRL research on this topic at TNO.