Generalization Ability of Deep Learning Algorithms Trained Using SEM Data for Objects Classification

Abstract

This paper proposes a workflow to efficiently determine the material of spherical objects and the location of the receiving antenna relative to their position in bi-static measurements using supervised learning techniques. From a single observation, we compare classification performances resulting from the application of several classifiers on different data types: the Ultra-Wide Band scattered field in time and frequency domains and pre-processed data from the Singularity Expansion Method (SEM). Indeed, the resonances extracted using the SEM are aspect independent and therefore, are used to discriminate the objects. As for the residues, they depend upon the aspect angle and can hence be exploited to determine the observation angle. We construct 3 datasets to assess which one yields the highest accuracy while using the simplest and fastest classifiers. Hence, 80% of each dataset is used for training and the remaining 20% are used for testing. In a further step, we test with sphere sizes and data with several noisy levels that were not in the training datasets. Although SEM is noise sensitive, associating a robust feature extraction technique (Vector Fitting) with a simple but reliable classifier is promising, particularly when generalizing to data not included in the training set.

Shape and orientation classification of objects based on their electromagnetic signatures using convolutional neural networks

“…Our first objective is to address the multi-class classification of objects with different geometries using convolutional neural network (CNNs). The choice of CNN was made considering previous studies that highlighted its efficiency compared with other classifiers applied to real data (DT, multiLayer perceptron (MLP), SVM) [22,23]. Classifying an object's shape, regardless of its size, is an interesting task that has not yet been addressed using SEM data.…”

Robust Target Classification Using UWB Sensing*

“…More recently, in [25] the authors tested different ML algorithms to classify PEC targets with simple shapes by using CNRs for synthetic data. In [26] the author extended the application of the workflow to sphere targets of different compositions in addition to PEC. Again, this work was tested solely on simulated data, using both noiseless and noisy signals.…”

A feasibility study of AI-based classification of concealed threats at mm-Waves

“…Another point to be raised and kept in mind in future work is the importance of the pre-processing of the measured data, which includes the antenna response suppression and to deal with the support response of the objects. Several techniques were previously demonstrated to overcome this [7,3] but we have decided to focus on the image processing ones, such as the segmentation with the Otsu method.…”