2023-01-012024-05-14https://scholars.lib.ntu.edu.tw/handle/123456789/660442雞肉為國內重要肉品來源之一，而有色肉雞特別為國人所鍾愛。隨著國人對肉品品質要求逐漸提升，把關破損、瘀傷等瑕疵雞隻之重要性亦隨之提高。傳統上，業者多透過以人力進行瑕疵屠體之篩選。然而，此方法不僅費時且缺乏效率，更可能在過程中發生人為疏漏。因此，本計畫擬針對國內有色肉雞之屠宰場生產線，藉由雞隻屠體攝影系統與深度學習模型，進行有色肉雞之屠體自動化瑕疵偵測。雞隻屠體攝影系統將由工業攝影機、嵌入式系統(樹莓派)、密封防水盒、不鏽鋼支架與紅外線光電對射開關組成，並將架設於有色肉雞屠宰場生產線中之分級設備之前，拍攝個別屠體之正背面影像。收集之影像將在進行影像標記後，進一步建立深度學習模型以自動偵測屠體瑕疵。屠體瑕疵偵測將分為二階段進行，先利用物件偵測模型區分出翅膀、胸部、背部、腳四種部位後，再將各別部位之影像，以深度學習模型判斷該其是否有瑕疵之情況。本計畫期望透過開發雞隻屠體攝影系統，取代傳統之人力篩選，以提高國內有色肉雞之生產效率，並輔助有色肉雞屠宰流程之管理。 Native chicken is one of the major sources of meat in Taiwan. Nowadays, the importance of picking out the flawed carcass is rising with costumers are becoming stricter in meat quality. Conventionally, the flawed carcasses of the native chicken were screened by necked-eye examination. However, inspect the flaws manually is time consuming, inefficient, and error-prone due to fatigue. Therefore, this project plans to employ a camera system and deep learning models to detect the flaw of the carcasses automatically in a native chicken production line of an abattoir. The camera system will be composed of an industrial camera, an embedded system (Raspberry Pi), watertight casings, stainless steel frames, and the infrared photoelectric beam detector. The camera system will be installed before the weight grading system for capturing the images of the front and back of the chicken carcasses. After annotated the collected images, an object detection model (ODM) and a flawed carcasses identification model (FCIM) will be trained for identifying flawed carcasses automatically. The ODM will be used for classifying four body parts of the carcasses, namely wings, breast, back, and legs. Then, the FCIM will be used for identifying whether the body parts of the carcass are damaged. This project expects to increase the effectiveness of native chicken butcher and supporting the management of the abattoir.深度學習;肉品分級;瑕疵辨識;Deep learning;Meat grading;Damaged identification