Performance Dependence on System Parameters in Millimeter-Wave Active Imaging Based on Complex-Valued Neural Networks to Classify Complex Texture

Millimeter wave exhibits relatively straight propagation as well as high penetration into dielectric materials, such as plastics, cloth, and paper. In security imaging, we use these features to discover weapons concealed under clothes. Self-organizing map (SOM), a type of neural networks, can map high-dimensional data on any dimension with unsupervised learning. It is utilized for clustering and visualization of high-dimensional data. Previously, we proposed a millimeter-wave imaging system for moving targets consisting of a 1-D array antenna, a parallel front end, and a complex-valued SOM to deal with complex texture. Experiments demonstrated its high performance in the visualization. In this paper, we investigate the dependence of the visualization performance on its configuration parameters as well as processing parameters. We reveal the effect of the modulation-frequency number and the window size. We also discuss the effective depth range for visualization and a tradeoff relationship between the measurement time and the visualization quality.