Automatic measurement of acidity from roasted coffee beans images using efficient deep learning
1
Issued Date
2022-11-01
Resource Type
ISSN
01458876
eISSN
17454530
Scopus ID
2-s2.0-85135252320
Journal Title
Journal of Food Process Engineering
Volume
45
Issue
11
Rights Holder(s)
SCOPUS
Bibliographic Citation
Journal of Food Process Engineering Vol.45 No.11 (2022)
Suggested Citation
Sajjacholapunt P., Supratak A., Tuarob S. Automatic measurement of acidity from roasted coffee beans images using efficient deep learning. Journal of Food Process Engineering Vol.45 No.11 (2022). doi:10.1111/jfpe.14147 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/84042
Title
Automatic measurement of acidity from roasted coffee beans images using efficient deep learning
Author(s)
Author's Affiliation
Other Contributor(s)
Abstract
Sourness is one of the basic yet essential tastes of coffee that is chemically composed of acids and quantitatively represented in the pH scale. Current tools for measuring the acidity level in roasted coffee beans, including traditional methods, require brewing sample coffee and probing the chemical components, limiting the applicability to end customers seeking to estimate the acidity level before choosing the right coffee beans to purchase. This paper proposes a novel approach to directly estimate the acidity levels from roasted coffee beans images by framing the problem into an image classification task, where a picture of roasted coffee beans is categorized into its appropriate pH range. As a result, end customers could simply estimate coffee beans' acidity levels by taking photos with conventional cameras. Multiple traditional machine learning and deep learning algorithms are validated for their ability to predict the correct acidity levels. The experiment results reveal that EfficientNet yields the best performance with an average F1 of 0.71 when trained with images from separate portable devices. Practical Applications: The research's findings could also be extended to applications in the coffee-industrial settings, such as automatically monitoring roasted coffee beans' quality from image and video streams. For end customers, the trade-off between efficacy and efficiency of the EfficientNet algorithm is also investigated, which sheds light on the implementation aspects of state-of-the-art deep learning models in portable devices such as smartphones or cameras. Such applications could prove to be a cost-effective and convenient solution for customers to quickly measure roasted coffee beans' sourness before deciding to purchase.