|
摘要
隨著卷積神經網路在圖像檢測領域的快速發展，近年來學界與業界致力於將卷積神經網路技術導入到自動化光學檢測的系統內，進行工業瑕疵檢測，邁向產線智慧化。然而，對於新的工件，即便已有過往之相似樣品的檢測模型，也會因樣品本質上的差異或取像光學架構的差異，而無法直接使用，檢測效果極差，故往往只能選擇對新貧料重新標記訓練，而訓練檢測模型通常需耍大量的瑕疵様本標籤資料，其獲取方式往往仰賴手動標記，極耗費時間與人力。
有鑑於此，本研究使用領域自適應的技術，在未對待測之新資料進行標籤的情況下建立瑕疵分類模型。其原理為利用事先己標記且與待測目標資料相似的資源資料、待檢測的無標記目標資料進行模型訓練，讓神經網路最終可以成功檢測目標資料。本研究以實際產線的不同木種之木皮瑕疵影像、不同色系之紡織瑕疵影像與公開之金屬表面瑕疵資料集為例，比較領域自適應神經網路在不同工業應用的檢測表現，並分析網路擷取之特徵進行模型優化，根據結果統整出領域自適應技術應用於工業瑕疵檢測的合適流程。
經過本硏究的實驗測試，將 ResNet50 作為特徵擷取器訓練領域自適應模型 DANN，並輔以熵調整 (Entropy
Conditioning)，能有效分類無標籤之瑕疵影像。相較於直接使用舊有相似資料之模型進行辦識，對於木皮類之瑕疵影像，分類準確率能從 52.96%
提升至 84.93%；對於布匹類之瑕疵影像，分類準確率能從 22.58%提至 73.75%；對於金屬表面之瑕疵影像，分類準確率能從31.13%提升至 95.58%。另外，若對特徵擷取的特徵層進行優化選擇，分類準確率能再有所提升。通過此流能快速訓綀出對無標籤新資料的辨識模型，有效節省人力與時間成本
關鍵字：智慧製造、人工智慧、深度學習、領域自適應、瑕疵檢測
Abstract
With the development of deep learning,
convolutional neural network has achieved outstanding performance in
the field of image detection. In recent years, the academia and
industry have been committed to introducing convolutional neural
network technology into the production line of automatic optical
inspection. However, due to the differences of data characteristics
and image acquisition methods, it is ineffective to recognize defects
on new target data with a former model trained by similar data. Thus,
engineers usually manually label the new target data and train a new
defect recognition model, which brings a lot of labor costs.
In
view of the above, in this research, the defect recognition model is
built by usingdomain adaptation, which is able to train a neural
network on a labeled similar source
dataset and secure a good
accuracy on the unlabeled new target dataset. Wood and textiledefect
images gathered from actual production lines and an open dataset of
metal surface
defect images, NEU-CLS, are used as the verification
data. Different kinds of domainadaptation models are compared. Also,
feature extraction layers are analyzed to optimizethe model. Finally,
a general process to train a defect recognition model using domain
adaptation is organized.
According to
the results, a classification model trained by the DANN domain
adaptation method with a ResNet50 backbone and entropy conditioning
algorithms is effective to recognize unlabeled defect images. For the
wood defect dataset, the accuracy increases to 84.93%. For the textile
defect dataset, the accuracy increases to 73.75%. For the metal
surface defect dataset, the accuracy increases to 95.58%. In addition,
the accuracy is further increased by choosing proper feature
extraction layers. By following this process, an effective defect
recognition model can be built without labeling new data. In other
words, time and labor costs on labeling new data can be significantly
reduced.
Keywords: Smart
Manufacturing, Artificial Intelligence, Deep Learning, Domain
Adaptation, Defect Inspection
|
