| 摘要: | 隨著科技的日新月異，許多技術逐漸應用於工業領域，帶來了第四次工業革命，亦稱為工業4.0。其中，人工智慧(Artificial Intelligence, AI)與擴增實境(Augmented Reality, AR)的結合，為工業帶來了許多的優勢。在現代工廠中，技術人員在組裝零件時，可以即時透過AR眼鏡來查看零件的辨識結果，這需要高度準確的模型來支持。然而，僅依賴於零件的外觀和特徵進行辨識，在面對外觀相似但尺寸有差異的零件時，模型容易產生混淆，導致模型準確度下降。此外，零件的不同角度、照明條件和背景等外在環境因素也會影響模型的準確度。為了解決這一問題，我們開發了一套系統，該系統要求在外在環境干擾較低的情況下，利用擁有深度感測器的HoloLens 2來取得單一零件的RGB影像及深度影像，深度影像能提供感測器與物體之間的距離資訊。然後利用這些影像並透過我們所提出的方法來找出能容納零件的最小矩形，進而估算出其實際寬度與長度，並將我們的方法與K-means聚類演算法進行比較，該演算法也能找出容納物體的最小矩形並進行尺寸估算。找出最小矩形的目的是為了能夠統一各種形狀的零件，提供一個標準化的尺寸參考，減少計算複雜度。最後，再使用我們所估算出的尺寸資訊來對模型的辨識結果進行額外的處理。實驗結果表明，我們所提出的方法在尺寸估算上明顯優於K-means聚類演算法，並證明了模型在尺寸資訊的輔助下，能有效提升其準確度。;With the rapid advancement of technology, many innovations have gradually been applied to the industrial sector, ushering in the Fourth Industrial Revolution, also known as Industry 4.0. Among these innovations, the combination of Artificial Intelligence (AI) and Augmented Reality (AR) has brought numerous advantages to the industry. In modern factories, technicians can use AR glasses to view real-time identification results of components during assembly, which requires highly accurate models. However, relying solely on the appearance and features of components can lead to confusion when dealing with components that look similar but differ in size, resulting in decreased model accuracy. Additionally, factors such as different angles, lighting conditions, and backgrounds can also affect the model′s accuracy. To address this issue, we developed a system that, under conditions of minimal environmental interference, uses the HoloLens 2 equipped with a depth sensor to capture RGB and depth images of the individual component. The depth images provide information on the distance between the sensor and the object. Using these images, our proposed method identifies the minimum rectangle that can enclose the component, estimates its actual width and length, and compares our method with the K-means clustering algorithm, which also identifies the minimum enclosing rectangle and estimates dimensions. The purpose of finding the minimum rectangle is to standardize the dimensions of components of various shapes, reducing computational complexity. Finally, we use the size information we estimated to further process the model′s recognition results. Experimental results show that our proposed method significantly outperforms the K-means clustering algorithm in size estimation, and demonstrates that the model can effectively improve its accuracy with the assistance of size information. |
