Band-target infrared imaging for plastic flake analysis via hyperspectral band selection and object recognition.

Achieving a circular plastics economy requires the high-purity identification of plastic flakes among diverse non-plastic contaminants in recycling streams. To this end, the accurate and rapid measurement of impurity levels in flake feedstock is crucial for the industry to ensure high-quality secondary raw materials. Short-wave infrared hyperspectral imaging (HSI) is well-suited to this task because structure-specific CH vibrational overtones provide chemical fingerprints, yet the large spectral dimensionality of HSI imposes substantial acquisition and computational demands that have limited its practical deployment. Here, we present a filter-wheel-based band-target imaging (BTI) framework in which the multispectral band set is rationally derived from spectroscopic and hyperspectral analysis. By relating the hydrogen distribution of five plastic monomer units and two contaminants to their second- and first-overtone absorption features in the 1100-1700 nm region, six diagnostic bands are selected to maximize inter-class spectral variance. The implemented BTI system, coupled with a deep learning-based object detector, reduces the per-frame acquisition time to under 2 s for the entire 2D region while achieving a mean average precision (mAP@0.5) of 91.8% across the seven target classes. Ultimately, this integrated approach demonstrates the potential for chemically specific identification of plastic flakes, offering a promising step toward the sustainable recovery of high-purity plastic resources.