食品瑕疵檢測

Abstract

一、 中文摘要 精緻農產品及智慧化生產為台灣農業及工業重點發展對象，透過自動化流水線，結合深度學習及計算機視覺。本專題開發出一套適用於工廠內的自動化胡蘿蔔辨識系統，具體來說我們模擬出一個食品業實際情況，也就是隨機放置及不同光線下的狀態。在這種極具挑戰的前提下，我們完成了精準辨識出胡蘿蔔位置及範圍，透過語義分割及物件偵測。這項工作中考慮的農產品是胡蘿蔔，胡蘿蔔為許多國家大量生產的農產品之一。透過徹底的實驗及分析，我們將對20條的胡蘿蔔進行檢測評估，總共蒐集到3120張照片，其語意分割結果平均wIoU為0.9899，物件偵測結果平均F1-score為0.9549。以其穩定及精確度達到足以應付流水線上自動化生產的標準。預計這項研究可以統合到自動化流水線分級分類系統中，可以協助未來自動化農業的發展。

二、 Abstract Delicate agricultural products and intelligent production are the key development targets of Taiwan's agriculture and industry. Through automated assembly lines, combined with deep learning and computer vision. This topic develops a set of automatic carrot identification system suitable for factories. Specifically, we simulate an actual situation in the food industry, that is, the state of random placement and different light conditions. Under this extremely challenging premise, we have completed the precise identification of the location and range of carrots, through semantic segmentation and object detection. The agricultural product considered in this work is carrot, which is one of the agricultural products produced in large quantities in many countries. Through thorough experiments and analysis, we will conduct detection and evaluation on 20 carrots. A total of 3120 photos have been collected. The average wIoU of semantic segmentation results is 0.9899, and the average F1-score of object detection results is 0.9549. With its stability and accuracy, it can meet the standard of automatic production on the assembly line. It is expected that this research can be integrated into the automated assembly line classification system, which can assist the development of automated agriculture in the future.