A combination of a convolutional neural network, which belongs to the deep learning research field, and support vector machines is presented as an efficient automatic target recognition system. Additional training methods that incorporate prior knowledge to the classifier and further improve its robustness against imaging errors and target variations are also presented. These methods generate artificial training data by elastic distortion and affine transformations that represent typical examples of image errors, like a changing range scale dependent on the depression angle or an incorrectly estimated aspect angle. With these examples presented to the classifier during the training, the system should become invariant against these variations and thus more robust. For the classification, the spotlight synthetic aperture radar images of the moving and stationary target acquisition and recognition database are used. Results are shown for the ten class database with a forced decision classification as well as with rejection class.
1 December 2019 in IEEE Aerospace and Electronic Systems Magazine
The article presents the cognitive radar architecture of Fraunhofer FHR based on a three-layer model of human cognitive performance. The approach is illustrated using examples for non-cooperative target identification and classification. On the skill based layer, a target-matched waveform design is presented and experimental results are shown. For the transition to the rule-based layer, convolutional neural networks are explained for the identification of air-targets and a novel auto-encoder for change detection is introduced. For rule-based behavior, a policy based decision making algorithm for NCTI waveform selection is explained, using CPOMDPs.
Recently an increased interest in deep learning for radar and particularly SAR ATR systems has been observed. Many authors proposed systems that outperform established classification systems on benchmark datasets like MSTAR. In this paper we present a new implementation of our recently proposed convolutional neural network classifier, which has a more flexible structure and at the same time less free parameters and thus a reduced training time. Furthermore, the training itself is improved through regularization techniques that improve the convergence properties of the network. Another feature of this new implementation is the dependency of the learning rate on the target class. With this feature the network can focus on classes that cause higher costs of misclassification.
7 March 2023 in IEEE Transactions on Aerospace and Electronic Systems
With the rapid development of Unpiloted Aerial Vehicles (UAVs), also known as drones, in recent years, the need for surveillance systems that are able to detect drones has grown as well. Radar is technology with the potential to fulfill this task and several previous publications show examples of radar detection and classification schemes. The purpose of this paper is related to the detection scheme used in these radars. Most surveillance systems use a background subtraction and a threshold to detect targets. This threshold often depends on a model of the radar noise and the background, which is imperfect by nature. The approach presented here uses a data driven machine learning algorithm that is trained with measured background profiles of the radar and is applied afterwards to the given background for target detection. This scheme can in general be applied to any detection problem in a fixed area, but is shown here with examples from measurements of drones and persons. The results show that the chosen approach gives better detection rates for low false alarm rates with real data than background subtraction.
Morphological Component Analysis is a technique to separate morphological different components from an image or signal. Morphological difference is in this case measured by the dictionary incoherence of the corresponding components. We propose to use a local discrete cosine transform to represent the periodic ground clutter and an undecimated wavelet transform to represent the piecewise smooth target. The parameters of the algorithm, like the total variation constraint, are determined automatically dependent on the contrast of the images. The decomposition is demonstrated with the spotlight SAR image chips of the MSTAR database and the found target images are used as input for a classification system to show the benefit of an increased generalization capability. As classifier we use the recently proposed combination of a convolutional neural network and support vector machines. Results are shown for forced decision classification as well as with rejection class.
29 April 2025 in IEEE Transactions on Aerospace and Electronic Systems
The detection and classification of radar emitters from the electromagnetic spectrum is an essential task for electronic support. This article introduces an innovative method that directly estimates the radar emitters from the spectrogram of received data in a single step, utilizing a convolutional neural network. The processing does not rely on the typically used sequence of pulse detection, pulse deinterleaving, and emitter identification and does not require the creation of pulse descriptor words. The approach can detect multiple interleaved emitters in a spectrogram and is, therefore, a multilabel classification task. The feasibility of this single-step approach is evaluated for a library of 50 emitters, which includes high parameter agility. The results show that the proposed approach provides a high detection probability of up to 98 % while maintaining a low false alarm rate of only 0.02%. In addition, we propose an explainable artificial intelligence technique based on occlusion that is well suited to radar emitter recognition in spectrogram data. A final evaluation of the trained neural network with test signals generated in the lab investigates the robustness to some real-world radio frequency effects.