| 摘要: | 在產線製造過程中，瑕疵分類可以有效檢測產品品質且利於人員修復產品，近年來雖然可以透過 AI 輔助瑕疵分類來大幅降低時間及人力成本，但一般分類模型應用於瑕疵分類上準確度往往差強人意，其原因在於瑕疵類別之間十分相似，使其成為了細粒度圖像分類問題，這使得模型需要能更精準地學習到更重要的瑕疵特徵才能避免模型被相似的類別混淆，還有在對新產品進行瑕疵分類時，會因為新、舊產品的背景顏色和亮度不同，而發生了 domain shift 問題，導致模型的效能會大幅下將。
雖然這些問題都有各自解決的方法，但在瑕疵分類中，模型需要能同時改善上述問題，因此我們提出了 Unsupervised Fine Grained Domain Adaptation Model combine with Critical Feature Alignment for Defect Images Classification (UFD)，此方法以細粒度圖像分
類模型為架構來強化對相似瑕疵類別之間的辨別力，並生產偽標籤使得模型透過訓練沒有標籤的新產品來增強模型泛化力，再提出了 critical feature alignment (CFA) 的方法來改善 domain shift 的問題，最後透過使用預訓練權重來增加訓練初期偽標籤的正確性使模型能夠順利訓練並縮短訓練時間。
在實驗中，我們使用非公開的面板瑕疵資料來進行實驗，而在真實場景的瑕疵分類實驗中，我們的方法在四個平均指標中都高出其他方法約 20% 至 30%。，這證實了我們的方法可以有效應用於瑕疵分類任務中，以達到降低人力與時間成本的目的。此外，我們還使用了兩個公開的資料集來進行實驗，目的在於證明我們方法能夠改善細粒度圖像分類問題以及 domain shift 問題，而我們的方法在實驗二與最好的方法相差不到 0.3%，實驗三我們的方法比第二好的方法提升約 2%，並且跟其餘的方法相比提升了
10% 以上。;In manufacturing, defect classification plays a crucial role in effectively assessing product quality and facilitating repair by personnel. In recent years, leveraging AI for defect classi fication has significantly reduced time and labor costs. However, the performance of general classification models in defect classification is unsatisfactory. This is primarily due to the high similarity among defect categories, which introduces challenges related to fine-grained image classification. To address this, models must precisely capture critical defect features to avoid confusion between similar categories. Furthermore, domain shift issues often arise when clas sifying defects in new products due to differences in background color and brightness between new and old products, leading to a substantial decline in model performance.
Although these challenges have individual solutions, defect classification models must ad dress them simultaneously. To this end, we propose the Unsupervised Fine-Grained Domain Adaptation Model with Critical Feature Alignment for Defect Image Classification (UFD). This method employs a fine-grained image classification architecture to enhance discrimination be tween similar defect categories. Then, we generate pseudo-labels to enable the model to train on unlabeled new product data, thereby improving generalization. In addition, we propose a Critical Feature Alignment (CFA) method to mitigate domain shift issues. Finally, by using pre-trained weights to increase the accuracy of pseudo-labels in the early stages of training, the model can be trained smoothly, and the training time can be shortened.
In the experiments, we utilized a proprietary panel defect dataset to evaluate our method. In real-world defect classification experiments, our approach outperformed other methods by approximately 20% to 30% across four averaged metrics. This demonstrates that our method can be effectively applied to defect classification tasks, achieving the goal of reducing labor and time costs. Additionally, we conducted experiments on two public datasets to validate the ef fectiveness of our method in addressing fine-grained image classification problems and domain shift issues. In Experiment 2, our method was within 0.3% of the best-performing approach, while in Experiment 3, our method outperformed the second-best approach by approximately 2% and achieved an improvement of over 10% compared to other methods. |
