國產針葉材表面缺點偵測研究

Abstract

木材中有從樹木生長到加工生產流程中可能出現的缺點，最主要為樹木側枝基部殘留於樹幹所形成的節，會大幅降低木材的性質。目前國內對於木材缺點的檢查多為人工進行，為了自動化木材生產流程，我們需要能夠自動偵測木材缺點的技術。 本研究採用faster region convolution neural network (Faster R-CNN)物件偵測模型，將木材表面影像輸入模型後，由模型輸出該影像中存在的缺點位置、尺寸以及種類資訊。影像資料來自於杉木(Cunninghamia lanceolata)、柳杉(Japonica cryptomeria)以及臺灣杉(Taiwania cryptomerioides)，皆為國內常見的人工林樹種，共計1,238張木材表面影像。訓練由8種不同CNN特徵提取骨架所構成之Faster R-CNN，在位置與尺寸的邊界框偵測上，Precision與Recall最高可達到93.18%以及92.23%，分別為Resnet50_FPN和Resnet18_FPN模型；在缺點分類準確率則是Resnet50_FPN有最高的87.06% Accuracy，同時可達到60.06%的mean average precision (mAP)。在偵測錯誤的案例中，側枝最外圍所造成的不正常纖維排列是缺點認定的模糊地帶，人工判定存在個人主觀成分導致不確定性。此外，生節在木材中自然出現的頻率遠高於其他種類的缺點，造成缺點分類的不平均。 本研究作為木材加工化的基礎，在木材表面缺點偵測上取得令人滿意的成果，未來能結合木材之目視分等或集成元加工等實際應用的自動化，提升木材產品的品質與穩定性。

There are certain defects associated with wood, primarily the presence of knots formed by the residual branches at the base of the tree, which significantly reduce the quality of the wood. Currently, the inspection of wood defects in the domestic market relies mostly on manual processes. To automate the wood production process, technique that can detect wood defects automatically is necessary. In this study, the Faster Region Convolutional Neural Network (Faster R-CNN) object detection model is adopted. By inputting surface images of wood into the model, it can output the location, size, and class of defects. The image data used in this study consisted of 1,238 images of commonly found plantation species in Taiwan, including Chinese fir (Cunninghamia lanceolata), Japanese Cedar (Japonica cryptomeria), and Taiwania (Taiwania cryptomerioides). The Faster R-CNN models was trained using 8 CNN feature extraction backbones, achieving the highest precision and recall of 93.18% and 92.23%, respectively, for bounding box detection, with the ResNet50_FPN and ResNet18_FPN models. The ResNet50_FPN model also achieved the highest accuracy of 87.06% for defect classification and a mean average precision (mAP) of 60.06%. In cases where errors occurred, the abnormal fiber arrangement caused by the outer periphery of knots created ambiguity in identification, leading to uncertainty due to subjective judgments. Additionally, the frequency of occurrence of live knots in wood is much higher than other types of defects, resulting in an imbalance in classification. This study has achieved satisfactory results in the detection of surface defects in wood, as a foundation for wood processing. In the future, integrating this automated technique with visual grading and processing of lumber could enhance the quality and consistency of wood products.