隨著智慧製造技術的迅速發展，深度學習於工業領域中扮演越來越關鍵的角色。然而，對於以接單生產為主的工廠而言，要事先取得真實零件拍攝大量的真實影像以訓練深度學習模型極為困難。為解決缺少訓練影像的問題，將採用3D模型檔產生合成影像作為替代資料。由於3D模型檔缺乏紋理資訊，導致其生成影像與真實影像存在差異，進而影響模型在真實場景中的辨識效能。本研究提出結合深度學習技術之工業零件分類與長度估算框架，採用二階段辨識流程，拍攝多角度之RGB與RGB-D影像，透過深度資訊估算目標零件長度，並將長度資訊作為輔助特徵，過濾分類模型預測結果，將不可能之類別排除，以提升細粒度分類準確度。且為解決工廠環境複雜因素，在推論時，整合圖像切割模型建立背景去除模組，自動化移除背景干擾，提升模型專注於零件特徵之能力並解決長度估算的背景條件限制。同時，考量工業零件之幾何特性差異，提出基於形狀的長度估算模組，以降低誤差並提升整體長度估算準確性。實驗結果驗證本系統框架在實際場域中的可應用性及具備良好的準確度性能，在Top1準確率可達到60.72%，Top5準確率可達到84.26%。;With the rapid advancement of smart manufacturing technologies, deep learning has become critical in the industrial domain. However, it is challenging for order-oriented factories to collect a large number of real-world images for training deep learning models. To address the issue of limited training data, 3D models are used to generate synthetic images as an alternative data. Due to the lack of texture information in most 3D models, resulting in a domain gap between synthetic and real-world images, which affects the recognition performance of the models in real-world scenarios. In this study, we propose deep learning-based framework for industrial components classification and length estimation, which adopts a two-stage recognition process by taking RGB and RGB-D images from multiple angles. Depth innformation is used to estimate the length of the target components, and the estimated length is further utilized as an auxiliary feature to filter the classification result, filtering improbable classes to improve fine-grained classification accuracy. In order to solve the complexity of the factory environment, this study integrates the segmentation model to build an background removal module, which automatically removes the background interference, improves the ability of the model to focus on the part features, and solves the limitation of the background conditions for length estima- tion. Meanwhile, the shape-based length estimation module is proposed to reduce the error and improve the overall accuracy of length estimation for the geometrical characteristics of 3D industrial components. Experiments results show that our propoosed method has the applicability of the system framework in real-world scenarios with good accuracy performance, achieving Top1 accuracy of 60.72% and Top5 accuracy of 84.26%.
