Classification of objects in polarimetric radar images using CNNs at 77 GHz

Due to their outstanding performance in image recognition, the application of convolutional neural networks (CNNs) for classification of polarimetric radar images is investigated in this paper. At first, several artificial radar targets that differ in physical dimensions and polarimetric properties are designed and constructed. These targets are mounted on a rotary table in an anechoic chamber and are rotated in front of a fully polarimetric radar operating at 77 GHz. For each rotation angle step, the targets are measured, where all measurements are taken at several distances. Then, a CNN is used to classify the measurement results into one of three superordinate classes of the constructed radar targets, each containing two different subordinate targets (sub-targets). Focus is lied on class decisions and not identification of specific sub-targets. A decision is taken for each single rotation angle step measurement and also for a majority vote on five consecutive rotation angle step measurements. Promising accuracy results demonstrate the validity of the approach but also reveal several open issues for future research.