水果成熟度檢測

Abstract

目的 傳統檢測方法通常是仰賴經驗老道的人肉眼觀察，或是使用一 些設備儀器(如硬度檢測儀、屈折度計等)，但上述儀器的檢測方式 大多數都會破壞到被檢測水果的外表完整性，直到近年來(西元 2021 年)才出現可以在不破壞水果外表完整性的情況下偵測其成熟 度:近紅外光譜儀，但由於其尚未商業化量產，尚無法幫助臺灣本地 番茄農民檢測番茄成熟度，因此我們決定使用圖像辨識來幫忙解決 這一難題。 過程及方法 本專題使用 YOLOv7 與 DETR 模型進行水果的成熟度辨識，使用 tomatOD 資料集訓練與測試模型，接著使用自行去番茄園拍攝的照 片資料集來依序進行成熟度區分、訓練集和測試集依照 8:2 比例分 配、標註各種成熟度出現比例。 下一步是進行模型訓練，分別使用 YOLOv7 與 DETR 來訓練，比 較兩邊的特色、訓練參數、損失值曲線、F1-score Confidence 曲 線、Precision Recall 曲線等，得出 YOLOv7 與 DETR 兩者的優劣。 結果 根據得到的評估指標，YOLOv7 除了 Precision 略差一點以外， 其他三項指標都明顯優於 DETR，所以我們認為 YOLOv7 在判斷番茄 成熟度上的成效較 DETR 優秀。若能結合物聯網進行遠端監測及使用 更多台灣本地番茄資料來訓練的話，有望造福台灣本地的番茄農民 可以更加便利的照顧與採收番茄。

Abstract Traditional detection methods usually rely on naked eye observation by experienced people, or use some equipment (such as hardness testers, refractometers, etc.). However, most of the above-mentioned detection methods will damage the appearance integrity of the fruits being tested. It was not until recent years (2021 AD) that a near-infrared spectrometer appeared that could detect the ripeness of the fruit without damaging its appearance integrity. However, because it has not yet been commercialized and mass-produced, it is not yet able to help local tomato farmers in Taiwan detect tomato ripeness. degree, so we decided to use image recognition to help solve this problem. Process and Methods This topic uses the YOLOv7 and DETR models to identify fruit maturity, uses the tomatOD data set to train and test the model, and then uses the photo data set taken by myself in the tomato garden to sequentially distinguish the maturity, training set and test set according to 8: 2 Proportional distribution and marking of the occurrence proportions of various maturity levels. The next step is to conduct model training, using YOLOv7 and DETR to train respectively. Compare the characteristics, training parameters, loss value curves, F1-score Confidence curves, Precision Recall curves, etc. of both sides to get the advantages and disadvantages of YOLOv7 and DETR. Conclusion According to the obtained evaluation indicators, except that Precision is slightly worse, the other three indicators of YOLOv7 are significantly better than DETR, so we believe that YOLOv7 is better than DETR in judging tomato maturity. If the Internet of Things can be combined with remote monitoring and more local tomato data in Taiwan can be used for training, it is expected to benefit local tomato farmers in Taiwan and make it easier to care for and harvest tomatoes.