本研究旨在詳細研究影像前處理如何影響導光板瑕疵檢測，並探討各種深度學習模型在不同的影像處理策略下的表現。導光板瑕疵檢測在確保產品品質上具有關鍵性的角色，然而，傳統的檢測演算法因為受到產品設計結構和製造過程變化的影響，無法有效地檢測出所有的瑕疵。為了解決這一問題，我們應用了最近在影像辨識領域取得突破的人工智慧技術。 我們分析了不同導光板結構的影像，並在三種不同的影像前處理策略下，評估了多種深度學習模型的預測效能。實驗結果顯示，隨著影像尺寸的增加，大部分的模型性能都有所提升。具體來說，Inception V3和EfficientNetB0在所有實驗中表現最為優秀，同時，使用Crop + Padding (Black) 或 Crop + Padding (Gray)的前處理策略也可以獲得較好的效能評估結果。 在機器學習模型方面，邏輯斯回歸（LR）顯示出短的訓練和測試時間以及良好的效能評估結果。此外，我們也發現不同的前處理策略會導致不同的訓練和測試時間，例如，在SqueezeNet和InceptionV3模型中，使用Crop + Padding策略可以縮短訓練和測試時間。 透過本研究，我們更深入地理解了影像前處理如何影響導光板瑕疵檢測，並評估了不同深度學習模型在不同影像處理策略下的效能。這些結果不僅有助於提升導光板瑕疵檢測的準確性和效率，也為未來相關領域的研究提供了重要的參考。 ;This study aims to thoroughly examine how image pre-processing impacts the detection of defects in light guide plates and investigate the performance of various deep learning models under different image processing strategies. The detection of defects in light guide plates plays a critical role in ensuring product quality. However, traditional detection algorithms are incapable of effectively detecting all defects due to variations in product design structures and manufacturing processes. To address this issue, we have applied recent breakthroughs in image recognition from the field of artificial intelligence. We analyzed images of different light guide plate structures and evaluated the predictive performance of multiple deep learning models under three different image pre-processing strategies. Experimental results indicate that as image size increases, the performance of most models improves. Specifically, Inception V3 and EfficientNetB0 demonstrate outstanding performance across all experiments. Additionally, the pre-processing strategies of Crop + Padding (Black) or Crop + Padding (Gray) also yield better performance evaluation results. In terms of machine learning models, logistic regression (LR) shows shorter training and testing times and exhibits good performance evaluation results. Furthermore, we found that different pre-processing strategies lead to varying training and testing times. For instance, the use of Crop + Padding strategy can shorten training and testing times in SqueezeNet and InceptionV3 models. Through this study, we gain a deeper understanding of how image pre-processing influences the detection of defects in light guide plates and evaluate the performance of various deep learning models under different image processing strategies. These results not only contribute to improving the accuracy and efficiency of defect detection in light guide plates, but also provide important references for future research in related fields.
