Refine
Document Type
Language
- English (18) (remove)
Has Fulltext
- no (18)
Is part of the Bibliography
- no (18)
Keywords
- Neural networks (7)
- Fluorescence (6)
- Fluorescence lifetime imaging (6)
- Phase shift (5)
- Microscopy (4)
- fluorescence (4)
- FD-FLIM (3)
- Laser sources (3)
- MLP (3)
- Bandpass filters (2)
Previous research on identifying plastic types and differentiating plastics from environmental material is promising by utilizing the specific fluorescence lifetime, but the evaluation still has to be automated. Therefore, an automated Gaussian analysis is developed for evaluating frequency-domain fluorescence lifetime images of plastics and environmental materials. Furthermore, we applied a “Multilayer Perceptron” and “Random Forest Classifier” to the data resulting from the Gaussian analysis of the frequency domain fluorescence lifetime imaging microscopy data. The classification results show high F1-scores, whereby the best “Multilayer Perceptron” and “Random Forest Classifier” achieved an F1-score of 90%. Thus, identifying and differentiating plastics and environmental materials is possible by applying a “Multilayer Perceptron” or “Random Forest Classifier” to the Gaussian-analyzed imaged fluorescence lifetime data.
Environmental pollution by plastics is an increasing problem. However, state-of-the-art methods have significant disadvantages in detecting and identifying plastics directly in an environmental matrix. In this study, we propose a blob detection algorithm in combination with a neural network for fast and automated identification of plastics and non-plastics in a single fluorescence lifetime image. Therefore an artificial environmental matrix is prepared that contains soil, grass, spruce and HDPE (high density polyethylene) particles. Several FD-FLIM (frequency domain fluorescence lifetime imaging microscopy) images are taken, and the detection algorithm and the neural network are applied. We successfully demonstrated the suitability of the thresholding algorithm and the binary classification of the HDPE particles directly in the environmental matrix.
Existing plastic analysis techniques such as Fourier transform infrared spectroscopy and Raman spectroscopy are problematic because samples must be anhydrous and identification can be hindered by additives. This article describes a new approach that has been successfully demonstrated in which plastics can be classified by neural networks that are trained, validated, and tested by frequency domain fluorescence lifetime imaging microscopy measurements.
The FD-FLIM technique has a high potential for automated post-consumer wood sorting. A problem of analysing post-consumer wood fluorescence properties is the uncertainty of the post-consumer wood category as the origin of the samples are unknown. In this study, the fluorescence properties of actual post-consumer wood is compared with prepared wood samples. The post-consumer wood samples display slightly different fluorescence intensities and lifetimes due to environmental influences and a higher sample diversity. For improved training of evaluation algorithms for post-consumer wood sorting, the prepared sample set should be extended or the post-consumer wood should be additionally analysed in the laboratory.
Development of a Neural Network for Automatic Classification of Post-Consumer Wood Using Rapid-FLIM
(2023)
The economic use of wood is a growing sector, not only because of the significant advantage of wood to retain CO 2 . It is crucial to increase the material recycling of wood in several lifecycles, but currently, there is no reliable post-consumer wood sorting technique in line. This research measures 365 post-consumer wood samples of classes A1-4 four times with the frequency domain fluorescence lifetime imaging microscopy (FD-FLIM) subset method rapid fluorescence lifetime imaging microscopy (Rapid-FLIM). The data is analyzed on their statistical features. Four neural networks based on Multilayer perceptron are then trained and tested with twelve statistical features extracted from the Rapid-FLIM images. The best model for this application contains the optimizer RMSprop, the activation function SELU and the loss function binary crossentropy. The best model of this structure could achieve a false positive ratio of 4.79 % over the ten folds.
The fast classification and identification of plastics presents a significant challenge. The study assesses the suitability of a Multilayer Perceptron to classify and identify commonly found plastic types using Rapid-FLIM, achieving an accuracy of 88.33%. @inproceedingsSchwarz:23, Increasing demand for wood products requires methods to determine its harvest origin and ensure sustainable and legal sourcing. In 15 out of 21 cases, the origin of Quercus robur was differentiable in FD-FLIM studies.
For sustainable handling of wood and high material utilization, post-consumer wood has to be sorted by post-consumer wood categories in Germany. The non-destructive FD-FLIM method could improve the wood recycling process by automatic classification of waste wood into post-consumer wood classes. Thus, the phase-dependent fluorescence decay time of treated wood is analysed in this study to obtain the impact of the treatment on the fluorescence decay time. It is shown that treated and untreated wood can be differentiated using the fluorescence decay time, although the state-of-the-art theoretical fluorescence decay time differs from the experimentally determined fluorescence decay time.