以語義分割檢測木皮表面瑕疵

Abstract

由於木皮表面的瑕疵種類涵蓋了很多不同的形狀及紋理同時又有複雜的背景，因此以影像處理的方法達成木皮表面瑕疵檢測是一個複雜的問題。自動化光學檢測引進到林木業產線來提升製程的效率並維持產品的品質有一段時間，然而現今深度學習技術的發展逐漸成熟已經超越了許多傳統的電腦視覺方法，並且嘗試取代以往所使用的自動化光學檢測技術。同時因為全捲積網路 (Fully Convolutional Networks)的提出，使捲積網路可以對圖像執行像素等級的分類以達到語義分割的目標。 本研究採用在公開資料集上表現較佳的DeepLab 網絡架構來執行木皮表面的瑕疵檢測，並與另一個網絡架構U-Net 比較預測的結果。在影像預處理上，採用影像切割去除多餘的背景部分可以提高訓練的瑕疵樣本比例，同時有效的提升訓練的預測結果。除此之外，使用重要樣本挖掘的方法 (Top K Percent Mining Method)在DeepLab 上更可以將整體的平均並交比 (Mean Intersection over Union)有效提升，改善後的召回率 (Recall Rate)更能達到70 %。實驗的結果顯示 DeepLab是適合用在木皮表面瑕疵檢測，並且在本研究建立的流程也同時可以適用在其他類型的工業檢測上。

Detection of wood sheet surface defects using image processing is a complicated problem in the forest industry as the image of the wood surface contains kinds of defects in different sizes, colors and textures. The automatic optical inspection (AOI) technology was introduced into the wood manufacturing line to improve production efficiency and maintain the quality of products for years. However, deep learning has achieved great success and outperformed traditional computer vision methods at present. With the development of fully convolutional networks (FCN), convolutional neural networks (CNN) can perform pixel-wise classification for image semantic segmentation. A well-performing segmentation architecture in open datasets, DeepLab is practiced in this thesis compared with the baseline model U-Net. The image preprocessing by cutting off the redundant background area can improve the training result. Also, by applying the top k percent mining method to DeepLab, the mean intersection over union (MIoU) can be boosted a lot, and the recall rate is up to 70 % with the improvement. The results show DeepLab is good at capturing contextual information and suitable for wood sheet surface defects detection. Furthermore, the defect detection process established in this thesis can be applied on other industrial detection besides wood sheet.