Identifying moldy peanut using hyperspectral imaging by correction of noisy labelling.

Abstract

Aflatoxin, a secondary metabolite synthesized by moldy peanuts, poses a significant and potentially fatal threat to human health. Hyperspectral imaging technology combined with supervised learning algorithms has become an essential method for rapid, non-destructive identification of moldy peanuts. However, it is impossible to measure the aflatoxin content of peanut kernel at pixel-level, weak labels is commonly phenomenon when initially labeling moldy regions within peanut hyperspectral images, which leads to an inevitable issue-noisy label. In this study, a label quality quantification framework which integrate confidence level and statistical testing (CL-ST) is proposed to find label errors of moldy peanut hyperspectral images. First, confidence level of initial artificial labeling is estimated using out-of-sample probability. Furthermore, label quality is assessed through statistical testing and ranked in descending order. Finally, the optimal noisy rate (NR) is determined based on moldy peanut kernel-scale identification performance, and the models with clean data as input is rebuilt to identify moldy peanuts. Experimental results show that CL-ST can reliably quantify label quality of hyperspectral images and effectively identify potential noisy labeled pixels. The rebuilt successive projection algorithm-extreme learning machine achieve optimal performance, improving overall accuracy and precision from 74.63 % and 66.21 % to 98.96 % and 97.09 %, respectively. Feature visualization analysis reveals that noisy labels are the primary cause of incorrect decision boundaries, although their impact on feature selection is limited. CL-ST does not require hyperparameters and can be combined with any model to quantify label quality, demonstrating significant potential in food quality assessment using hyperspectral images. The source code of CL-ST will be available at https://github.com/yuandeshuai/CL-ST.

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.