Abstract

With the goal of improving machine learning approaches in inverse scattering, we provide an experimental data set collected with a 2D near-field microwave imaging system. Machine learning approaches often train solely on synthetic data, and one of the reasons for this is that no experimentally-derived public data set exists. The imaging system consists of 24 antennas surrounding the imaging region, connected via a switch to a vector network analyzer. The data set contains over 1000 full Scattering parameter scans of five targets at numerous positions from 3-5 GHz.

Instructions:

The data set consists of JSON files. Each of the five targets was moved to a particular position and a full scan of S11, S21, S12, and S22 data were taken for that position. Each File of S-parameter data is separate. Files can be read in with any common JSON reader. There were 24 antennas in the system.

Each JSON file contains a 'meta' field that gives the start/stp frequency, VNA settings, dates and time of data collection, and the target name.

The position of the center of the target is given in mm. The expected permittivity of the target is also provided for dielectric targets. For circular (cylindrical) targets, the radius of the target is in the target name.

Each day of data collection, we also collected incident field data, were the imaging chamber was left empty. This empty chamber data is labelled 'incident field data' and in each JSON file we have a 'suggested_inc_field_dir' (or suggested incident field directory) that points to the empty/incident chamber data. In addition, for calibration purposes, we also scanned a 3.5 inch metallic cylinder at the start of each day, and we point to that data with the suggested calibration directory field.

All frequencies are listed in the 'freq' field of the JSON file.

The data for each position of the switch is contained in the 'data.txry' field. For example, accessing the 'data.t1r6' field will return the data for when the switch was set to transmitter number 1, receiver number 6 (the data is then returned for all frequencies).

The data is stored in real/imaginary format.

Funding Agency:

Natural Sciences and Engineering Research Council of Canada