利用葉片影像與卷積神經網路辨識殼斗科與樟科之物種

Abstract

森林是陸域生態系重要的一環，不但孕育豐富的動植物，更能提供木材、調節氣候，對人類社會與環境帶來諸多效益。為了更有效率應用及管理這些資源，準確的物種辨識是不可或缺的步驟。殼斗科與樟科是台灣中低海拔的優勢樹種，因為台灣豐富的地形與氣候變化分化出許多台灣特有種。現今的植物物種辨識依靠比對遺傳標記的相似性，然而該方法昂貴並耗時。近年來卷積神經網路 (convolutional neural networks, CNNs) 廣泛被應用於各種複雜的機器視覺任務。因此本研究利用深度卷積神經網路自動辨識35種殼斗科與樟科物種，其中葉片影像由彩色平板掃描機所取得。本研究之自動辨識模型使用三種不同的深度卷積神經網路架構，包含：DenseNet-121、MobileNet V2與Xception。研究中卷積神經網路模型之準確率最高可達99.396%，而在圖像顯示卡 (Graphics Processing Unit, GPU) 最快之辨識速度可達17.1毫秒∕影像。

Forests contain abundant resources and provide benefits to human societies, such as edible fruits, medicinal substances, and woods for construction. Species identification is an essential step for the management and utilization of the resources. Nowadays, the species identification relies on the examining the similarity of genetic markers. The approach is, however, time consuming, laborious, and costly. Fagaceae and Lauraceae are two woody plant families with high species richness that dominate the low and middle altitude regions in Taiwan. This paper proposed a machine vision approach for identifying the species of families Fagaceae and Lauraceae using leaf images. Leaf specimens of 35 Fagaceae and Lauraceae species were collected. The images of the leaves were acquired using flatbed scanners. Deep convolutional neural networks (DCNN) of three architectures, namely DenseNet-121, MobileNet V2, and Xception, were next trained to identify the species. Among all, Xception reached the highest mean accuracy of 99.396%, and MobileNetV2 required the least mean test time of 17.1 ms per image using an GPU.