MWIR-4-Plastic

MWIR-4-Plastic The Identification of Complex End-of-Life Industrial Plastic using Mid-wave Infrared Hyperspectral Imaging and Machine Learning Description The MWIR-4-Plastic dataset is a multimodal RGB and hyperspectral imaging dataset comprising 13 hyperspectral scenes of black/dark engineering polymers, curated to support the identification and classification of carbon-black-loaded plastics from shredded end-of-life vehicle (ELV) recycling streams. The scans are collected in the VNIR and SWIR ranges using a Specim FENIX (Spectral Imaging Ltd, Oulu, Finland) and in the MWIR range using a Specim FX50 hyperspectral camera, in addition to high-resolution RGB images using the JAI LT-400 CL RGB camera. The data consists of 758 bands between [400-5300] nm. The HSI data contains 450 bands from the FENIX camera in the VNIR and SWIR range [400-2500] nm, and 308 bands from the FX50 camera in the MWIR range [2700-5300] nm. Classes of interest: Styrene - PA - PC - PP The dataset spans 13 scanning sessions organized under four class-labelled folders (ABS/Styrene, PA, PC, PP) plus one mixed-class folder (mix) used for multi-object validation scenes. Each class folder contains scenes acquired under different sample configurations (class and particle size). The mixed scenes represent an end-of-life vehicle sorting product enriched in polymer particles before type separation. The remaining eleven scenes represent each polymer concentrate obtained after sensor-based sorting. The complete dataset comprises more than 15,000 shredded automotive particles, most of which are black. Particles were distributed across two nominal size fractions with object diameter ranging approximately in the first one from 5 mm to 11 mm and from 5 mm to 30 mm, with particle counts ranging from 500 (single-class scenes) to 2000 (mixed-class scenes) particles. Data acquisition was performed at the Helmholtz Institute Freiberg for Resource Technology using a specialized SisuRock hyperspectral scanner (Specim, Oulu, Finland) operating at a fixed working distance. More information about the acquisition, preprocessing, processing, and experiementation can be found in the publication: The Identification of Complex End-of-Life Industrial Plastic using Mid-wave Infrared Hyperspectral Imaging and Machine Learning Data Format HSI Data: Each hyperspectral cube is stored as a paired data/header set (“FENIX.dat”/“FENIX.hdr” and “FX50.dat”/“FX50.hdr”; the mix scenes use “FENIX.img”/“FX50.img” in place of “.dat”). RGB Data: High-resolution RGB image, stored as “RGB.dat”/“RGB.hdr”, with a rendered preview “RGB.png” (and, for some scenes, “RGB_false_color2.png” or “RGB_True_color2.png”). Preview Images: “FENIX.png” and “FX50.png” are false-color previews of each cube; “FX_false_color.png” is an additional MWIR false-color rendering. Ground Truth (semantic): “gt.png” give the pixel-wise class labels for the scene. “dTree_gt_ .npy” stores the corresponding manually verified ground-truth labels as a NumPy array (naming follows the class and sample configuration, e.g. “dTree_gt_PA11mm.npy”; mix scenes use “dTree_gt1.npy”/“dTree_gt2.npy”). Semantic mask: “semantic_mask2.png” Instance arrays: “instance2.pkl” contains the instance-level segmentation masks (object-wise); mix/1 additionally includes “instance.png” as a rendered preview. Folder Organization MWIR-4-Plastic/├── ABS_Styrene/ (2 scenes: EBS12_11mm_5mm_Styrole2, EBS22_SLF_Styrole2)├── PA/ (3 scenes: EBS12_11mm_5mm_PA2, EBS12_30mm_5mm_PA2, EBS22_SLF_PA2)├── PC/ (3 scenes: EBS12_11mm_5mm_PC2, EBS12_30mm_5mm_PC2, EBS22_SLF_PC2)├── PP/ (3 scenes: EBS12_11mm_5mm_PP2, EBS12_30mm_5mm_PP2, EBS22_SLF_PP2)└── mix/ (2 scenes: 1, 2 — mixed-class validation scenes) Each scene subfolder (e.g. PA/EBS12_11mm_5mm_PA2/) contains: FENIX.dat / FENIX.hdr / FENIX.png FX50.dat / FX50.hdr / FX50.png RGB.dat / RGB.hdr / RGB.png FX_false_color.png gt.png semantic_mask2.png instance2.pkl dTree_gt_ .npy Note: the mix/ scenes follow the same file set but use FENIX.img/FX50.img instead of FENIX.dat/FX50.dat, and dTree_gt1.npy/dTree_gt2.npy instead of the class-based naming convention used elsewhere. Code Repository To facilitate reading and working with the data, in addition to the adoption and reproducing of the results, the Python codes of ML and DL training and inference are available on the GitHub repository:https://github.com/hifexplo/MWIR-4-Plastic Citation Please cite the following article when using the dataset: MWIR-4-Plastic: The Identification of Complex End-of-Life Industrial Plastic using Mid-wave Infrared Hyperspectral Imaging and Machine Learning