基于K-means聚類算法的烘烤煙葉圖像分割研究

摘要 目前我國煙葉烘烤過程主要依賴人工監(jiān)測(cè)，存在主觀性、模糊性和高成本等問題，使用機(jī)器視覺方法對(duì)烘烤過程煙葉質(zhì)量變化進(jìn)行實(shí)時(shí)監(jiān)測(cè)與判斷的研究逐漸增多，實(shí)時(shí)監(jiān)測(cè)需建立在高效且準(zhǔn)確的烘烤煙葉圖像分割之上，因此烘烤煙葉圖像分割的研究變得尤其重要。提出了基于K-means聚類算法的烘烤煙葉圖像分割方法，首先讀取圖像并將RGB轉(zhuǎn)換為CYMK顏色空間，然后提取CYMK顏色空間下的K通道灰度化圖像，再對(duì)此單通道圖像進(jìn)行聚類，根據(jù)聚類中心確定圖像分割閾值，最后利用圖像處理方法對(duì)圖像進(jìn)行分割。研究比較了K-means、模糊C均值聚類（FCM）和高斯混合聚類（GMM）3種聚類方法，結(jié)果表明K-means算法的像素準(zhǔn)確率為97.8%、交并比為96.43%、Dice系數(shù)為98.2%，均優(yōu)于其他2種方法。K-means算法能夠更好地提取烤煙的煙葉輪廓，去除冗余信息，使得分割結(jié)果更清晰。

關(guān)鍵詞 煙葉烘烤；圖像分割；K-means；閾值

中圖分類號(hào) S-058 文獻(xiàn)標(biāo)識(shí)碼 A 文章編號(hào) 0517-6611（2024）19-0232-06

doi：10.3969/j.issn.0517-6611.2024.19.047

開放科學(xué)（資源服務(wù)）標(biāo)識(shí)碼（OSID）：

Research on Curing Tobacco Image Segmentation Based on K-means Clustering Algorithm

ZHOU Ren-hu，XI Jia-xin，DING Yi-shu et al

（Chuxiong Company of Yunnan Provincial Tobacco Company，Chuxiong，Yunnan 675000）

Abstract At present，the tobacco baking process in China mainly relies on manual monitoring，which has problems of subjectivity，fuzziness and high cost.Research on using machine vision methods to monitor and judge real-time changes in tobacco quality during the baking process is gradually increasing.Real time monitoring needs to be based on efficient and accurate segmentation of roasted tobacco leaf images，so the research on segmentation of roasted tobacco leaf images has become particularly important.A segmentation method for roasted tobacco leaf images based on K-means clustering algorithm was proposed.Firstly，the image was read and RGB was converted to the CYMK color space.Then，the grayscale image of the K-channel in the CYMK color space was extracted.We clustered the single channel image again，determined the image segmentation threshold based on the cluster center，and finally used image processing methods to segment the image.We compared three clustering methods of K-means，fuzzy C-means clustering （FCM） and Gaussian mixture clustering （GMM）.The results showed that the pixel accuracy of the K-means algorithm was 97.8%，the intersection to union ratio was 96.43%，and the Dice coefficient was 98.2%，all of which were better than the other two methods.The K-means algorithm could better extract the contour of tobacco leaves，remove redundant information and make the segmentation results clearer.

Key words Tobacco curing;Image segmentation;K-means;Threshold value

基金項(xiàng)目 中國煙草總公司云南省公司科技項(xiàng)目（2022530000241034）。

作者簡介 周任虎（1977—），男，云南楚雄人，農(nóng)藝師，碩士，從事煙葉烘烤技術(shù)研究與應(yīng)用。

收稿日期 2023-08-03

在我國的煙草產(chǎn)業(yè)中，烤煙作為一種重要的經(jīng)濟(jì)農(nóng)作物，對(duì)卷煙生產(chǎn)有著重要的影響。烘烤是煙葉生產(chǎn)過程中非常關(guān)鍵的環(huán)節(jié)工藝，它直接關(guān)系到烤煙葉的品質(zhì)和產(chǎn)量^［1］。目前，我國煙葉烘烤過程仍主要依靠專業(yè)人員進(jìn)行人工監(jiān)測(cè)，根據(jù)經(jīng)驗(yàn)和行業(yè)標(biāo)準(zhǔn)進(jìn)行烘烤程度的判斷，但這種方法具有主觀性和模糊性，并且人工成本較高^［2］。……