Evaluation of Deep Learning-Based Segmentation Methods for Industrial Burner Flames

Julius Großkopf,Markus Vogelbacher,Patrick Waibel,Jörg Matthes

doi:10.3390/en14061716

Julius Großkopf, Markus Vogelbacher + Show 2 more

Open Access

https://doi.org/10.3390/en14061716

Copy DOI

Journal: Energies	Publication Date: Mar 19, 2021
Citations: 4	License type: CC BY 4.0

Affiliation: Karlsruhe Institute of Technology

Abstract

The energetic usage of fuels from renewable sources or waste material is associated with controlled combustion processes with industrial burner equipment. For the observation of such processes, camera systems are increasingly being used. With additional completion by an appropriate image processing system, camera observation of controlled combustion can be used for closed-loop process control giving leverage for optimization and more efficient usage of fuels. A key element of a camera-based control system is the robust segmentation of each burners flame. However, flame instance segmentation in an industrial environment imposes specific problems for image processing, such as overlapping flames, blurry object borders, occlusion, and irregular image content. In this research, we investigate the capability of a deep learning approach for the instance segmentation of industrial burner flames based on example image data from a special waste incineration plant. We evaluate the segmentation quality and robustness in challenging situations with several convolutional neural networks and demonstrate that a deep learning-based approach is capable of producing satisfying results for instance segmentation in an industrial environment.

Highlights

The segmentation of flames is motivated by applications that require geometric information about flames in image or video data, see, e.g., in [1], where the authors use growing fire regions as a criterion to identify dangerous fire in outdoor and indoor scenery
We evaluated several recent convolutional neural networks (CNNs) architectures for the task of instance segmentation on combustion images in an industrial plant
We obtained experimental instance segmentation results with an IoUf ≥ 0.7 of both flames on more than 90% of our test images

Summary

Introduction

The segmentation of flames is motivated by applications that require geometric information about flames in image or video data, see, e.g., in [1], where the authors use growing fire regions as a criterion to identify dangerous fire in outdoor and indoor scenery. [2] use segmentations of burner flame images to derive a geometric stability measure for combustion processes. The procedure essentially require precise and reliable flame segmentation. In this context, the search for suitable segmentation methods for flame regions is a contribution to combustion process control that opens the door for more capable and energy-efficient combustion processes. Other methods channel multiple features into a more complex classification process such as SVM [7], Bayes classifier [4,6], neural network [3], or fuzzy logic classifier [5]

Objectives

Methods

Findings

Conclusion