| dc.description.abstract | 當今晶圓製造業中，自動光學檢查(Automatic optical inspection, AOI)已經成
為關鍵技術，然而，由於晶圓的大小和機台支援的影響，儘管自動光學檢查技
術在圖片檢驗領域取得了顯著進展，但仍然無法完全滿足客戶需求，導致需要
大量人工進行檢驗，以確保檢查的準確性。這種人工檢驗的過程是耗時和費用
高昂的，也會導致成本的上升。因此，透過使用人工智慧(Artificial Intelligence,
AI)技術，自動化進行檢查降低成本，是本研究的主要方向。
本文探討了一種基於深度學習的自動光學檢查檢驗後人工辨識輔助檢驗優
化方法，以滿足客戶需求，減少人工檢驗，提高檢驗準確率並實現快速部署和
檢驗。為解決這一問題，我們首先通過已建立的人工檢驗平台對資料進行標記。
在標記資料平衡的情況下，我們採用 合作式注意力網絡(Convolution and
Attention Network, CoAtNet)方法進行 AI 圖片分類。而在標記資料不平衡的情況
下，我們使用對抗自編碼器(Adversarial Autoencoder, AAE)與基於 GAN 的高效異
常檢測(Efficient GAN-Based Anomaly Detection, EGAN)方法生成模型。接著，我
們搭配閥值的調整進行 AI 圖片異常判斷，實現異常檢測。
通過進行一系列實驗，我們證明結合了合作式注意力網絡與對抗自編碼器
生成 Golden Sample 的方法可以顯著提高自動光學檢查檢驗後人工檢驗的準確率，
減少人工檢驗數量，並滿足客戶需求。本研究為圖片檢驗提供了一個有效且實
用的解決方案，具有廣泛的應用前景。 | zh_TW |
| dc.description.abstract | In today′s semiconductor wafer manufacturing industry, Automatic Optical
Inspection (AOI) has become a key technology. However, due to the size of wafers and
the constraints of equipment capabilities, even though AOI technology has made
significant progress in image inspection, it still cannot fully meet customer
requirements. This results in a high dependency on manual inspection to ensure
accuracy. This manual inspection process is both time-consuming and costly, leading
to increased production costs. Hence, automating the inspection process to reduce costs
using Artificial Intelligence (AI) technology is the main direction of this research.
This paper discusses an optimization method for post-AOI manual identification
assistance based on deep learning. The aim is to meet customer needs, reduce manual
inspections, improve inspection accuracy, and achieve rapid deployment and inspection.
To address this challenge, we first labeled data through an established manual
inspection platform. With balanced labeled data, we adopted the Convolution and
Attention Network (CoAtNet) approach for AI image classification. In cases of
unbalanced labeled data, we employed the Adversarial Autoencoder (AAE) and
Efficient GAN-Based Anomaly Detection methods for model generation. We then
paired this with threshold adjustments for AI image anomaly determination, realizing
anomaly detection.
Through a series of experiments, we demonstrated that combining the cooperative
attention network with the adversarial autoencoder to generate Golden Samples can
significantly enhance the accuracy of post-AOI manual inspections, reduce the number
of manual inspections, and meet customer needs. This research provides an effective
and practical solution for image inspection and has a broad application prospect. | en_US |