2019-08-012024-05-18https://scholars.lib.ntu.edu.tw/handle/123456789/704262摘要：殼斗科和樟科植物常見於台灣常綠闊葉林中，是生態系中重要的優勢樹種以及生產者。殼斗科和樟科植物不僅常用於建築與家具用材，其萃取物也被證實擁有抗菌、抗發炎等有效成分，成為近年極具價值的經濟樹種。正確的物種辨識是森林資源利用之始，物種的鑑定錯誤常會造成林產生產上的經濟損失。傳統上，殼斗科和樟科植物的鑑定仰賴長期研究該類植物的專家，以肉眼觀察葉片特徵進行辨識。然而，對專家而言，殼斗科和樟科植物的辨識是相當大的挑戰，因其葉片顏色、形狀、紋理和葉脈等外觀形態具有相當大的種內變異，需累積大量的觀察時間與經驗才得已準確辨識。本計畫將使用機器學習和深度學習技術，利用葉片影像辨識樟科和殼斗科物種。本計畫為期三年。第一年，以殼斗科物種的葉片影像為辨識對象，同時也蒐集樟科植物葉片影像。影像前處理包含影像切割與影像擷取的自動化程式開發，特徵包含顏色、形狀、紋理和葉脈，並以支持向量機和稀疏表示分類器對量化的特徵進行物種辨識。第二年，以殼斗科物種的葉片影像為辨識對象，同時亦繼續殼斗科植物葉片影像蒐集。同樣以顏色、形狀、紋理和葉脈為特徵，以支持向量機和稀疏表示分類器進行樟科物種辨識。第三年，以深度學習分類器之卷積神經網絡和生成式對抗網路進行殼斗科和樟科物種辨識，以每種2500張葉片影像為目標，持續蒐集這兩科物種之葉片影像。此年度亦進行反卷積網路將卷積神經網絡和生成式對抗網路訓練之卷積層視覺化，以提供深度學習分類器中辨識不同物種之葉片特徵。經由一些目標物種的葉子圖像分析，初步結果顯示，我們所提出的樹種辨識方法可以達到合理的精確度。本計畫預期透過計畫的執行訓練機器學習和深度學習的專家，將執行本計劃的知識與經驗應用於社會與工業，以建立台灣成為人工智慧之島。最後，本計劃也預期開發一套電腦視覺輔助的工具於殼斗科和樟科植物的辨識，輔助林業從業人員、森林保育員和環境教育講師等於野外植物辨識的工作，同時也供國人了解台灣森林植物多樣性與資源的契機。<br> Abstract: Plants of families Lauraceae and Fagaceae dominate the subtropical evergreen forest in Taiwan. The derivatives of the plants provide significant economic benefits. Correctly identifying the species in the field is critical as the first step to utilize the biomaterials. Identifying the species also helps scientists and ecologists for the purposes of forest resource conservation. Leaves of plants in families Lauraceae and Fagaceae show variation in color, shape, and texture. Conventionally, the identification of the species is conducted using naked-eye inspection with leaf characteristics. However, expert training is time consuming and expensive. Moreover, within certain genus, there exists a high degree of interspecific similarity in leaf morphology. The morphological differences between the leaves are so subtle that it is challenging, even for experts, to distinguish the species. Thus, in this project, we propose to identify Lauraceae and Fagaceae species using leaf images, machine learning, and deep learning approaches. This is a 3-year project. In year one, we will focus on leaf specimen collection for species in family Fagaceae. Leaf specimens of species in family Lauraceae will also be collected if it is admissible. Leaf images will be acquired using flat-bed scanners. Subsequently, leaf traits, including color, shape, texture, and venation, will be quantified. Machine learning classifiers, such as support vector machine and sparse representation classification classifiers, will be developed to identify the species using the traits as inputs. In year two, we will focus on leaf specimen collection for species in family Lauraceae. Leaf specimens of species in family Fagaceae will also be continued. Machine learning classifiers will be developed to identify the species. In year three, leaf specimens of species in both families will be continued. Deep learners, including convolutional neural networks and generative adversarial networks, will be developed to discriminate the species. Deconvolutional networks will be applied to visualize the learned leaf features that are significant for species identification. We have collected some leaf images of the target species. Preliminary results showed that the proposed methods could reached reasonable accuracies. By conducting this project, we expect to cultivate experts in machine learning and deep learning. The experts will subsequently propagate the learned knowledge to the society and industry, eventually increasing the compatibility of Taiwan. We also expect to develop a computer-aided tool for identifying plants in families Lauraceae and Fagaceae. The tool can assist the forest explorers in plant identification. It also can provide information to the general public when they visit the nature.深度學習卷積神經網路生成式對抗網路機器學習支持向量機稀疏表示分類器葉片影像樟科殼斗科Deep learningconvolutional neural networksgenerative adversarial networksmachine learningsupport vector machinesparse representation classificationleaf imagesLauraceaeFagaceae