利用手機影像與深度神經網路辨識木材

Abstract

臺灣擁有豐富的森林資源，其中木材是最具經濟價值且廣泛使用的森林產物。不同木材物種在價格與應用上存在顯著差異，因此準確有效率的辨別木材物種對於有效的木材資源利用至關重要。傳統木材辨識方法主要包括顯微分析與宏觀觀察。前者仰賴專家進行組織切片染色並於顯微鏡下進行觀察，儘管辨識準確率高仍需耗費大量時間進行樣本製備，難以應用於現場即時辨識；後者依賴人員對氣味、質地與色澤等感官特徵進行判斷，此方法高度仰賴人員經驗，主觀性高且缺乏一致性。相較之下，基於影像辨識的方法提供一種高效率且低成本的替代方案。近期，部分研究開始嘗試以智慧型手機拍攝之影像搭配深度學習技術進行木材辨識；然而，此類方法仍面臨跨裝置色彩差異及異常影像（如非目標物種及瑕疵影像）導致的誤判問題。為解決上述問題，本研究針對臺灣常見的41種木材，提出一套結合智慧型手機影像與深度學習技術之自動化木材辨識系統。研究共蒐集134,370張掃描影像用於模型訓練與初步測試，並另外蒐集15,743張手機影像用於模型微調及測試。本研究所提出的辨識系統包含三個核心階段：（i）色彩校正模組，有效降低不同拍攝裝置之色彩差異；（ii）異常影像檢測模型；用以排除非目標物種與瑕疵影像；（iii）基於視覺轉換器（Vision Transformer）的木材辨識模型。在色彩校正方面，不同裝置間的平均ΔE*色差經校正後由13.59降至3.02，色差降低幅度達77.2%；異常影像檢測模型準確率達 98.31%，而物種辨識模型於智慧型手機影像上的準確率達 94.71%。此外，整個系統透過雲端平台部署，並以 API 連結 LINE 聊天機器人介面，提供即時現地的木材辨識服務。本研究所開發之系統突破傳統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時間。

Taiwan possesses abundant forest resources, among which wood is the most economically valuable and widely utilized forest product. Given the substantial variation in price and application across wood species, accurate and efficient species identification is essential for effective resource utilization. Conventional wood identification methods primarily include microscopic analysis and macroscopic observation. Microscopic analysis relies on experts to prepare histologically stained sections and examine anatomical structures under a microscope. While this approach yields high accuracy, it is time-consuming due to the complexity of sample preparation, making it unsuitable for real-time, on-site identification. In contrast, macroscopic observation depends on field personnel assessing sensory characteristics such as odor, texture, and color. This method is highly subjective, relies heavily on individual experience, and often leads to inconsistent results. In contrast, image-based identification approaches offer an efficient and objective alternative. Recent studies have explored smartphone-based wood identification using deep learning; however, these methods still encounter challenges arising from cross-device color variation and the presence of anomalous images (e.g., non-target species or defective samples). To address these challenges, this study proposes an automated wood identification system specifically developed for 41 wood species commonly found in Taiwan, comprising 27 hardwoods and 14 softwoods. A total of 134,370 scanner images were collected for model training and testing, and 15,743 smartphone images used for model fine-tuning and testing. The proposed pipeline comprises three core components: (i) a color calibration module to effectively mitigate cross-device color discrepancies, (ii) an out-of-distribution (OOD) detection model to exclude non-target or defective images, and (iii) a wood identification model based on Vision Transformer architecture. In terms of results, after color calibration, the mean ΔE* color difference between devices decreased from 13.59 to 3.02. The OOD detector achieved 98.31% accuracy, and the wood identification model achieved 94.71% accuracy on smartphone images. The complete system is deployed on a cloud platform and integrated with a LINE chatbot via API, offering real-time, and user-friendly, identification services. This approach further reduces the reliance on expert labor and significantly shortens the time required for accurate identification.