深度學習陰影去除之咖啡粒徑分析系統

Abstract

隨著精品咖啡的迅速發展，家庭沖煮咖啡已經變得非常普遍，不同的沖煮手法和沖煮參數會造就不一樣的咖啡表現方式，水粉比、水溫與咖啡粒徑是國際咖啡協會認定的最主要的沖煮變因。時至今日粒徑的測量仍然缺乏方便且準確的方式進行測量，影響咖啡師在實務上對於粒徑的控制問題。 為了解決這問題，本研貢獻兩個重要的創新。首先，開發一款粒徑檢測軟體 (CPASR System)，只需一張白紙，上方撲滿咖啡粉，使用手機拍攝圖像並傳輸到 電腦中，透過本研究的軟體，即可得到粒徑分佈圖，使得咖啡師能夠有效地控制 粒徑，從而影響咖啡風味的表現。 其次，本研究另一個重要貢獻是創建一個咖啡粉顆粒陰影與無陰影資料集，並公開分享給廣大研究者使用，這個資料集可供未來從事咖啡顆粒影像相關研究者使用，透過這個資料集研究者們可以更深入地研究，並進一步改進咖啡粉陰影去除的方法和技術。 而在獲得咖啡粉圖像時，由於拍攝角度問題和咖啡粉本身立體特性，可能會有陰影的產生，在分析上會產生誤差，本研究陰影去除模型:CPASR Net 結合深度學習與傳統影像處理方法，來消除咖啡粉圖像中的陰影，最後通過多種粒徑估計方式，來實現粒徑的準確計算。 總體而言，本研究不僅開發了一個實用的咖啡粒徑檢測軟體，讓咖啡師能夠更好地了解咖啡粉粒徑的分佈，也貢獻了一個咖啡顆粒影像的資料集，促進了咖啡相關研究的發展。同時，通過陰影去除模型的引入，我們提高了粒徑分析的準確性，使得咖啡師和研究者們在咖啡品質改進和研究方面更具信心和可行性。

With the rapid development of specialty coffee, home brewing has become increasingly common. Different brewing techniques and parameters result in diverse coffee profiles, where water-to-coffee ratio, water temperature, and coffee particle size are recognized as the primary brewing variables by the International Coffee Association. However, to this day, measuring coffee particle size still lacks a convenient and accurate method, affecting baristas' ability to control particle size effectively in practice. To address this issue, this research introduces two significant innovations. Firstly, we have developed a particle size detection software called CPASR System. By simply placing a sheet of paper covered with coffee grounds and capturing an image with a smartphone, users can transfer the image to a computer and obtain a particle size distribution chart using our software. This empowers baristas to control particle size more effectively, influencing the expression of coffee flavors. Secondly, another important contribution of this study is the creation of a dataset containing coffee powder particles with and without shadows. This dataset is made publicly available for researchers in the field of coffee particle imaging. Researchers can utilize this dataset to delve deeper into the subject, improving methods and techniques for removing coffee powder shadows. When capturing images of coffee grounds, the presence of shadows due to angles and the three-dimensional nature of the coffee powder may introduce errors during analysis. To tackle this issue, our research introduces the shadow removal model, CPASR Net, which combines deep learning with traditional image processing methods to eliminate shadows from coffee powder images. The accurate calculation of particle size is achieved through multiple particle size estimation techniques. In conclusion, this research not only developed a practical coffee particle size detection software, enabling baristas to better understand the distribution of coffee particle sizes, but also contributed a dataset of coffee particle images, fostering advancements in coffee-related research. Furthermore, with the introduction of the shadow removal model, we have enhanced the accuracy of particle size analysis, instilling greater confidence and feasibility for baristas and researchers in improving coffee quality and conducting further studies.