基於葉片影像特徵的植物物種自動辨識研究

Research on Plant Species Automatic Identification via Leaf Image Recognition

盧琬臻(Wan-Jen Lu)

指導教授 ： 黃乾綱

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2017年)

https://doi.org/10.6342/NTU201701219

摘要

地球上各種樹木植物是最重要的自然資源之一，提供人類棲所與光合作用產生氧氣，全世界的植物物種有成千上萬種，該怎麼分類透過植物影像分類物種、依照外觀特徵辨識該物種是個很重要的議題，未來若能透過藉由行動裝置拍攝的植物影像達到物種辨識的效果，將會是個很方便又快速的途徑去認識植物。
以最常見且一年四季皆可取得的植物葉片而言，葉片上包含了許多不同的特徵，例如葉片邊緣鋸齒、葉脈走向、葉尖與葉基形狀、葉片長寬比例、葉片顏色、葉片表面質地與材質等等，可以藉由不同的葉片特徵來辨別植物物種。以往的研究方法使用一組葉片特徵值組合來描述所有影像中的植物葉片，但並非每項特徵在所有植物葉片上皆是重要特徵。有鑒於舊方法不能完整並精確描述各物種葉片，故本研究提出全新的兩階段分類方法，第一階段分類用以分類葉片形狀，藉由觀察所有常見的葉片形狀，將植物葉片依照形狀區分四大類形狀：長條形、掌形、三角與心形、圓與橢圓形，此四種類別外觀差異最大。第二階段的分類是植物物種辨識，於每一形狀類別各自使用不盡相同的特徵值以區分植物物種，除了採用葉形特徵之外，還考慮了葉片鋸齒、葉片表面反光程度等能描述更多葉片細節的特徵值。
實驗證明針對不同形狀的葉形類別使用了不同組合的特徵值進行影像辨識，讓每個形狀類別皆能使用到最合適的特徵值，確實有效的降低辨識錯誤率。從研究結果可以確實觀察出圓與橢圓形的葉片形狀有較高的辨識難度，相對的其他三種類的葉片形狀分類錯誤率極低，整體植物物種辨識錯誤率僅 6.93%。於文章末討論辨識錯誤的葉片中造成辨識效果不佳的原因並討論解法。

關鍵字

植物物種辨識 ； 特徵值擷取 ； 影像處理 ； 植物葉片分析 ； 分類

並列摘要

A variety of plant species on the earth is one of the most important natural resources, plants could be a shelter for human and do photosynthesis to produce oxygen, there are thousands of plant species around the world, how to classify so many species by leaf image? Using the appearance, flower, and leaf of a plant is a very important issue. If we can use the plant images taken by a mobile device to identify plant species in the future, it will be a convenient and quick way to know plants.
Leaf is the most common organ of a plant, it can be obtained all year, the leaf contain many different features, such as leaf edge, veins, shape, leaf aspect ratio, leaf color and leaf surface texture, etc., people can identify a wide variety of plant species by there features. Previous studies used a set of leaf features to describe the leaf of all plant, but not every feature was an important feature on all plant. So this study propose a whole new approach by operate a two-stage classification. The first stage is a leaf shape classification. Observing all the common shape of the blade, there are four distinguish shape class: long, palmate, triangle and heart, round and oval. The second stage of classification is the recognition of plant species. In addition to the leaf shape features, the leaf sawtooth, the reflection degree on the leaf surface, the features ​​which can describe the details are used to identify plant species. In each shape class, the different features ​​are applied to different shape class.
This study shows that each shape class use the most suitable combination of features ​​will exactly has lower error rate of plant species recognition. It can be observed from the results that the round and oval shape class has a higher identification difficulty, while the other three classes of leaf shape has a very low error rate of classification. the overall of species recognition, error rate is only 6.93%. In the end of this study, we discuss the reasons for the wrong recognition.

並列關鍵字

plant recognition ； feature extraction ； Image processing ； leaf shape analysis ； classification

參考文獻

30. 陳加容, 基於手持移動裝置之室內空間文字影像擷取. 臺灣大學工程科學及海洋工程學研究所學位論文, 2016: p. 1-60.

2. Hati, S. and G. Sajeevan, Plant Recognition from Leaf Image through Artificial Neural Network. International Journal of Computer Applications, 2013. 62(17).

4. Bruno, O.M., et al., Fractal dimension applied to plant identification. Information Sciences, 2008. 178(12): p. 2722-2733.

5. Intaravanne, Y. and S. Sumriddetchkajorn. BaiKhao (rice leaf) app: a mobile device-based application in analyzing the color level of the rice leaf for nitrogen estimation. in Proc. SPIE. 2012.

11. Guru, D., Y.S. Kumar, and S. Manjunath, Textural features in flower classification. Mathematical and Computer Modelling, 2011. 54(3): p. 1030-1036.