Automated chemical imaging identification of illegal drugs in correctional facilities mail

The current opioid epidemic represents a significant health and security threat. This epidemic has affected correctional facilities with an increased smuggling of illicit drugs concealed in envelopes, letters, greeting cards, and business cards to inmates. Short‐wave infrared chemical imaging sensors are being successfully applied to the automated, high confidence detection of drugs concealed in prison mail. Once detected, end users have a need to confirm the detection and to identify the specific drug being detected. The challenge for identification is that the spectral signature of the concealed drug is often convolved with the spectral signature of the piece of mail (substrate) in which the drug is concealed. This paper presents a method to remove the substrate signal from the substrate/drug mixture signal followed by a set of 3 spectral identification methods. The substrate signature is estimated by a region of interest that is spatially local to the detection, and linear unmixing uses this substrate signature to calculate the residual spectra in the detection pixels. These residual spectra represent the isolated drug spectra, and they are compared with a spectral drug library via Euclidean distance, target factor analysis, and adaptive cosine estimator methods. This methodology was applied to a set of 116 positive‐ and negative‐control samples spanning a range of drugs and concealment methods with the result that 90 of the 104 positive‐control samples were identified correctly (86.5%) and 0 of the 12 negative‐control samples were incorrectly identified as a drug in the library (0%). This paper presents a method to identify illicit drugs concealed in mail articles at correctional facilities using short‐wave infrared chemical imaging sensors. Linear unmixing is used to separate the drug signature from the substrate signature. The resulting spectra are compared with a spectral drug library via Euclidean distance, target factor analysis, and adaptive cosine estimator methods. The methodology was applied to 116 samples with identification rates of 86.5% (positive samples) and 0% (negative samples).