Lukas J, Kolb S, Heinbuch J, Willig L, Plankenbühler T, Müller D, Karl J (2023)
Publication Type: Journal article
Publication year: 2023
Book Volume: 341
Article Number: 127705
DOI: 10.1016/j.fuel.2023.127705
Due to the increased use of inconsistent low-grade biogenic solid fuels and simultaneously stricter regulatory limitations for biomass combustion, there is a rising demand for suitable methods for an online characterization of fuels to allow for optimal plant operation. Aim of this work is to evaluate two camera-based machine vision approaches using a conventional feature extraction method based classification and a deep learning approach with a convolutional neural network (CNN) to distinguish six types of biogenic solid fuels and 20 fuel mixtures thereof. A comparison of four machine-learning algorithms applied to classify the samples based on all extracted color and texture features reaches a prediction accuracy of 93.5 %. An evaluation of the feature performance of single features shows that the selected color features (color thresholds and histogram-based features) are more relevant for distinguishing the fuels than the textural features (Haralick features and features extracted from the frequency domain). Despite the small size of the image dataset, also the CNN achieves a good prediction accuracy of 73.7 % for the given classification task. An increase of the number of images by fragmentation leads to a slightly increasing prediction accuracy for the deep learning approach while the accuracy of the conventional features learning approach decreases. Both approaches are suitable to distinguish six typical biogenic solid fuels and achieve high accuracies for the classification of 26 fuels and mixtures. While the feature learning is more accurate for the mentioned classification task, the CNN does not require prior feature extraction and would benefit from a larger dataset. These are promising results for an implementation as an online fuel monitoring in various applications, with further development in the robustness and extensions by means of a broader range of fuel mixtures being required.
APA:
Lukas, J., Kolb, S., Heinbuch, J., Willig, L., Plankenbühler, T., Müller, D., & Karl, J. (2023). Image-based biomass characterization: Comparison of conventional image processing and a deep learning approach. Fuel, 341. https://doi.org/10.1016/j.fuel.2023.127705
MLA:
Lukas, Johannes, et al. "Image-based biomass characterization: Comparison of conventional image processing and a deep learning approach." Fuel 341 (2023).
BibTeX: Download