Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Coded-Aperture Computational Millimeter-Wave Image Classifier Using Convolutional Neural Network

: Sharma, Rahul; Hussung, Raphael; Keil, Andreas; Friederich, Fabian; Fromenteze, Thomas; Khalily, Mohsen; Deka, Bhabesh; Fucso, Vincent; Yurduseven, Okan

Fulltext ()

IEEE access 9 (2021), pp.119830-119844
ISSN: 2169-3536
Journal Article, Electronic Publication
Fraunhofer ITWM ()
millimeter-wave; imaging radar; Computational Imaging; neural networks; image classification; coded-aperture

A millimeter-wave (mmW) classifier system applied to images synthesized from a coded-aperture based computational imaging (CI) radar is presented. A developed physical model of a CI system is used to generate the image dataset for the classification algorithm. A convolutional neural network (CNN) is integrated with the physical model and trained using the dataset comprising of synthesized mmW images obtained directly from the developed CI physical model. A ${k}$ -fold cross validation technique is applied during the training process to validate the classification model. The coded-aperture CI concept enables image reconstruction from a significantly reduced number of back-scattered measurements by facilitating physical layer compression. This physical layer compression can substantially simplify the data acquisition layer of imaging radars, which is realized using only two channels in this article. The integration of the classification algorithm with the CI numerical model is particularly important in enabling the training step to be carried out using relevant system metrics and without the necessity for experimental data. Leveraging the CI numerical model generated data, training step for the classification algorithm is achieved in real-time while also confirming that the numerically trained CI classifier offers high accuracy with both simulated and experimental data. The classifier integrated physical model also enables performance analysis of the classification algorithm to be carried out as a function of key system metrics such as signal-to-noise (SNR) level, ensuring a complete understanding of the classification accuracy under different operating conditions. The trained CI system is tested with synthesized mmW images from the physical model and a classification accuracy of 89% is achieved. The proposed model is also verified using experimental data validating the fidelity of the developed CI integrated classifier system. A classification latency of 3.8 ms per frame is achieved, paving the way for real-time automated threat detection (ATD) for security-screening applications.