The data and findings discussed here come from the cited paper and do not represent test results from VISPEK equipment.
Does waste plastic identification necessarily require a dense continuous spectrum? This study attempts to differentiate between common household plastics using a small number of visible and near-infrared discrete channels, comparing various machine learning methods. The uneven recognition performance of different polymers provides a concrete reference for evaluating the screening use of low-cost sensors.
Summary
This study uses discrete channel reflection sensors to collect common household plastics, compares dimensionality reduction and multiple classification models, and evaluates them through cross-validation. The results show that the difficulty of distinguishing each polymer is different, and the overall performance is still limited; color, surface and mixing state may change the signal, and the collected and recorded quantities cannot be regarded as independent physical quantities or universal sorting scales.
(AI-generated conceptual illustration for context; it is not data from the paper, a photograph of the research equipment, or a VISPEK product diagram.)
Why It Matters
Use a small number of spectral channels to pre-classify materials first, adding a layer of quantifiable material clues to recycling equipment.
Methods and Reference Measurements
The experiment uses the AS7265x three-chip sensor to read reflection signals in 18 discrete bands of 410–940 nm. The lighting and sample positions were fixed, and 200 records were collected for each of the six types of household plastics, for a total of 1,200 records. The data is dimensionally reduced by optional PCA or LDA, 10 classification methods such as random forest, support vector machine, and neural network are compared, and the recognition performance is evaluated through cross-validation.
Wavelengths and Acquisition Parameters
| Band/parameter | Original meaning and usage instructions |
|---|---|
| 410–940 nm | The original text collection or actual modeling scope of this case |
Key Findings
The average recognition accuracy of CNN is 72.50%, and that of MLP is 70.25%; the highest PS category is 83.5%, and the lowest PET is 66%. These are results on controlled experimental data and are not universal sorting accuracy for all waste plastics.
Limitations and Interpretation
Color, surface condition, and plastic mixes can affect the signal; 1,200 records do not equate to 1,200 physical objects from independent sources.
Band and model boundaries: This example uses original measurements or independent verification branches that lie entirely within 250–1700 nm.
Image copyright belongs to the paper’s original authors. License: CC BY 4.0.
Source Paper
Martinez-Hernandez U, West G, Assaf T. (2024). Low-Cost Recognition of Plastic Waste Using Deep Learning and a Multi-Spectral Near-Infrared Sensor. Sensors (Basel, Switzerland) 24(9). https://doi.org/10.3390/s24092821
Contact Us
Whether portable equipment can be used for similar detection needs to be evaluated based on samples and experimental conditions, and the specific effect still needs to be verified through actual testing. Welcome to contact VISPEK to exchange your sample and testing needs, jointly evaluate experimental plans, and explore the feasibility of using our company's equipment for actual measurement and verification.
- Email: contact@vispek.cn
- Website: https://www.vispek.cn
VISPEK products shown for illustration; these were not the devices used in the cited study.