摘要
咖啡早已是人們生活的一部分，但要產出一杯好咖啡的過程是非常繁瑣複雜的，為了確保所烘焙出來的咖啡品質符合需求，咖啡瑕疵豆剔除一定得在咖啡生豆烘焙前完成，然而，瑕疵豆剔除流程通常是經由人工手挑完成，這是一個極為耗費時間的必要流程，如果可以有效改進當前的方法，它將降低將一杯好咖啡端上餐桌的最終成本。因此，本研究希望藉由深度學習物件偵測技術來降低人力成本，物件檢測技術可以對目標進行分類並從圖像中標定其位置，利用該技術與硬體機構及電控元件有效結合，咖啡瑕疵豆的剔除將可被自動化。本研究結果顯示在5種分類的訓練資料，訓練25,000次的模型中呈現較好的效果，平均準確度達到.54，且在好咖啡豆的分類辨識上平均準確度也達.80。因此，本研究已完成該模型與電腦視覺結合的設計，能有效快速的辨識瑕疵豆及其位置，未來搭配自動咖啡豆剔除的機構設計，將有助於開發低成本的瑕疵豆自動剔除設備。
關鍵詞：深度學習、物件偵測、咖啡瑕疵豆
Abstract
Coffee has become a part of people’s lives, but good coffee production is very complicated. In order to ensure the quality of roasted coffee, defective coffee beans must be removed before roasting. However, the traditional manual removal of defective beans is a time-consuming and laborious process. If the current method could be improved, it would reduce the cost of the final process of getting a cup of coffee to the table. Therefore, this research hopes to use deep learning object detection technology to reduce labor costs. Object Detection technology can classify the target and find the position from the image. If it can be effectively integrated with hardware equipment, it would be able to automatically remove defective coffee beans. The research results show five types of training data, among which the model with 25,000 training steps can achieve .54 mean average precision, and the average precision can achieve as high as .80 in the good coffee categories. Moreover, this model has been combined with computer vision to assist users in identifying, locating, and removing defective coffee beans, thereby improving the overall picking speed. In the future, it can be combined with hardware architecture to complete automated picking.
Keyword: Deep Learning, Object Detection, Defective Beans