利用卷積神經網路及木材橫切面影像辨識木材物種

Abstract

在臺灣豐富的森林資源中，木材擁有最高的經濟價值，而正確的物種辨識為木材利用之始。傳統上，木材的物種辨識仰賴專家以組織染色將樣本染色於顯微鏡下進行辨識，既使辨識準確率高仍需要製備樣本的時間，以及仰賴專家的植物學知識與經驗。近年分子標記等方法被運用於辨識木材物種，儘管能夠準確辨識物種，此方法仍需要大量的時間、勞力成本、以及昂貴的設備。與之相比，影像辨識的方法則為相對高效率、低成本的物種辨識策略。因此，本研究提出使用深度學習中的卷積神經網路(Convolutional Neural Network)及一系列的資料處理步驟、搭配木材未染色之莖部橫切面樣本影像，辨識闊葉樹及針葉樹等41種台灣常見的商業木材之物種。本研究蒐集了大約三千片的木材標本影像並提出一個包含三步驟的資料處理流程：其一之物種分類模型以木材橫切面影像為輸入並輸出預測之物種及相對應的信心分數，其二之誤差控制模型用以剃除信心分數不足之輸入樣本，其三之可信度校正模型調整信心分數以符合準確率，以上三個模型都以掃描影像進行訓練並以手機影像測試。本研究亦開發了一款智慧型手機應用程式，包含圖形使用界面並與雲端辨識模型整合，進而提供使用者以攜帶裝置即時上傳影像進行木材物種之辨識。在結果方面，物種分類模型在訓練、驗證以及測試數據集上分別取得了94.50% ± 1.23%、89.71% ± 1.38%以及 74.24% ± 3.10%的辨識準確率；誤差辨識模型在拒絕0.321%、15.04%和48.36%的樣本下能有效提升0.45%、3.51%和13.72%的準確率；可信度校正則將期望校正誤差(Expected Calibration Error)從0.0826降低至0.0437。本研究提出之方法具備完整的系統架構且可全自動執行，結果則顯示本研究所提出之資料處理流程已能輔助第一線林業從業人員及海關人員查驗木材物種，透過本系統之開發與佈建將能消除以往木材辨識僅能在實驗室內進行之限制，並進一步降低木材辨識所需之專家人力與所需時間。

Forest resources in Taiwan are abundant. Among all forest resources, wood is the one that has the most economic values. Identification of wood species is conducted prior to the application and utilization of the materials. Conventionally, the species of wood specimens are identified through a series of histological treatment and microscopic observation. The process is still manual, time consuming, and it largely relies on experts’ experience. In recent decades, molecular marker-based methods have been frequently used for species identification with high accuracy; however, these methods are labor intensive, time consuming, and require high-end equipment investments. Image-based approaches, by contrast, are effective and efficient. Thus, this study proposed a solution of applying the convolutional neural network (CNN) model and a series of data processing methods to the wood species identification using the cross-section image collected from 41 commercial wood species in Taiwan. The images of approximately three thousand wood specimens were collected. A data processing pipeline containing three steps was proposed. The step of species classification model takes a cross-section image as input and output a classification result as well as a confidence score. The step of error rate control model ferried out the images that were not confident enough. The step of reliability calibration adjusts the confidence score to match the actual observed probability. All three models were trained using the images in the image database. A mobile application was developed to provide the service of wood species identification through a graphical user interface. The trained species classifier reached the training, validation, and test accuracy of 94.50% ± 1.23% (mean ± standard deviation), 89.71% ± 1.38%, and 74.24% ± 3.10%, respectively. The error rate control model improved the accuracy for 0.45%, 3.51%, and 13.73% while rejecting 0.32%, 15.04%, and 48.36% of the input images. Reliability calibration reduced the expected calibration error (ECE) to 0.0437, which indicates the identification model were well-calibrated. The proposed solution is complete and fully automatic. The results of the proposed pipeline indicate that the solution can assist first-line forestry workers and custom stuff in the task of wood identification. With the development and deployment of the proposed system, it eliminated the limitation of the wood identification that has to be conducted in laboratory. Thus, it can further reduce the time and effort of experienced experts to identify wood species.