- 學位論文
視覺特徵用於大規模學習:以晶圓圖與音樂曲風分類為例
Visual Features for Large-scale Learning: Case Studies on Wafer Map and Music Genre Classification
摘要
隨著大規模資料集日益容易取得，也讓大規模學習在學界與業界受到關注。其中，智慧型手機的龐大需求帶動了相關產業鏈的蓬勃發展。對於上游的半導體製造業而言，如何持續提升晶圓良率是一項重要的議題，而晶圓圖錯誤類型辨識是機器視覺裡的一個應用，其可將晶圓自動分類，以協助工程師尋找錯誤並加速良率提升。對於下游的APP產業而言，由於線上音樂串流服務的需求持續成長，也讓音樂曲風分類益發重要，其為機器聽覺裡的一個應用，可幫助音樂檔案的管理及推薦。然而，上述這兩項大規模學習的應用仍然缺乏精簡且具有鑑別力的特徵。與過去的研究不同，我們針對晶圓圖和歌曲分別設計視覺特徵。為了驗證系統的效能，我們建立了一個世界最大的公開晶圓資料集(WM-811K)，也使用世界最大的音樂曲風評測資料集(MASD)來驗證，實驗結果顯示我們提出的視覺特徵均可有效提升辨識率。此外，在晶圓圖錯誤類型辨識上，我們的方法已經實際在晶圓廠上線，而在音樂曲風分類上，我們的方法也獲得了MIREX 2011至2013年曲風分類競賽的冠軍，這均說明我們方法的穩健性。
並列摘要
Increased availability of large-scale datasets has attracted increased academic and industrial attention to large-scale learning. Concurrently, huge growth in demand for smart phones has had a commensurate impact on related industries such as wafer manufacturing and mobile application industries. In the wafer manufacturing industry, increased demand has driven efforts to increase wafer production capacity, in part by reducing failure rates. Wafer map failure pattern recognition (WMFPR), an application of machine vision, can be used to automatically classify wafers, thus assisting engineers in identifying root causes of failure and thus increasing wafer yield. In the mobile application industry, increased demand for online music distribution has driven interest in music genre classification (MGC), which is an application of machine hearing, can facilitate music organization and music recommendation for online music services. However, reduced yet discriminative feature representations are still needed for these two large-scale learning applications. By contrast to conventional approaches, we consider an alternate approach for designing visual features for WMFPR and MGC. To validate system performance, we collected the world's largest public wafer map dataset (WM-811k) for WMFPR, and applied the world's largest benchmark dataset (MASD) for MGC. Experimental results show that the proposed visual features can considerably improve recognition rates. Furthermore, TSMC has adopted the proposed WMFPR method, while the proposed MGC method won the MIREX music genre classification contests from 2011 to 2013, indicating the robustness of the proposed methods.