The Magic of Vision: Understanding What Happens in the Process
The first paper provides effective deep learning algorithms accompanied by comparative studies to support real-time condition monitoring for a specialized microwave drying process for porous foams being taken place in a confined chamber. The tools provided give its users a capability to gain visually-based insights and understanding for specific processes. We verify that our state-of-the-art deep learning techniques based on infrared (IR) images significantly benefit condition monitoring, providing an increase in fault finding accuracy over conventional methods. Nevertheless, we note that transfer learning and deep residual network techniques do not yield increased performance over normal convolutional neural networks in our case.
After a drying process, there will be some outputted images which are reconstructed by sensor data, such as microwave tomography (MWT) sensor. Hence, how to make users visually judge the success of the process by referring to the outputted MWT images becomes the core task. The second paper proposes an automatic segmentation algorithm named MWTS-KM to visualize the desired low moisture areas of the foam used in the whole process on the MWT images, effectively enhance users'understanding of tomographic image data. We also prove its performance is superior to two other preeminent methods through a comparative study.
To better boost human comprehension among the reconstructed MWT image, a colormap deisgn research based on the same segmentation task as in the second paper is fully elaborated in the third and the fourth papers. A quantitative evaluation implemented in the third paper shows that different colormaps can influence the task accuracy in MWT related analytics, and that schemes autumn, virids, and parula can provide the best performance. As the full extension of the third paper, the fourth paper introduces a systematic crowdsourced study, verifying our prior hypothesis that the colormaps triggering affect in the positiveexciting quadrant in the valence-arousal model are able to facilitate more precise visual comprehension in the context of MWT than the other three quadrants. Interestingly, we also discover the counter-finding that colormaps resulting in affect in the negative-calm quadrant are undesirable. A synthetic colormap design guideline is brought up to benefit domain related users.
In the end, we re-emphasize the importance of making humans beneficial in every context. Also, we start walking down the future path of focusing on humancentered machine learning(HCML), which is an emerging subfield of computer science which combines theexpertise of data-driven ML with the domain knowledge of HCI. This novel interdisciplinary research field is being explored to support developing the real-time industrial decision-support system.
Chalmers, Data- och informationsteknik, Interaktionsdesign (Chalmers)
Automated microwave tomography (Mwt) image segmentation: State-of-the-art implementation and evaluation
Journal of WSCG,; Vol. 2020(2020)p. 126-136
Artikel i vetenskaplig tidskrift
Condition Monitoring for Confined Industrial Process Based on Infrared Images by Using Deep Neural Network and Variants
ACM International Conference Proceeding Series,; (2020)p. 99-106
Paper i proceeding
Automatic image segmentation for microwave tomography (MWT): From implementation to comparative evaluation
ACM International Conference Proceeding Series,; Vol. 2019(2019)
Paper i proceeding
Smart tomographic sensors for advanced industrial process control (TOMOCON)
Europeiska kommissionen (EU), 2017-09-01 -- 2021-08-31.
Data- och informationsvetenskap
Chalmers tekniska högskola
CSE Jupiter 473
Opponent: Professor Timo Ropinski. Ulm University, Germany