利用卷積神經網路偵測有色肉雞屠體之瑕疵

Abstract

雞肉被廣泛認為是最常見的肉類之一，同時也在經濟上扮演著重要角色。為了提供消費者高品質的雞肉，必須在雞隻屠宰過程中挑選出有瑕疵之屠體。台灣有色肉雞屬於國內特有品種，相較於白肉雞，有色肉雞品種多樣性高，在外觀顏色與尺寸上都有相當大的差異，由於這些差異，目前市面上用於白肉雞之自動屠體瑕疵檢測系統並不適用於台灣有色肉雞。現行方法採用人力方式來挑選台灣有色肉雞屠體中的瑕疵部分，然而這種方法耗時且效率低下。因此，本研究旨在使用卷積神經網絡來實現對瑕疵屠體的自動辨識。為此，我們開發了一套影像模組，該模組安裝在屠宰產線上，用於收集屠體影像；並提出了兩階段的檢測方法來自動檢測有瑕疵之屠體。在第一階段，使用身體部位分割模型（BSM）來偵測雞隻屠體影像中的翅膀、胸部、背部和腳等部位。隨後，在第二階段，使用了瑕疵分類模型（DCM），分類出具有瑕疵之身體部位。在實驗結果中，BSM在辨識雞隻屠體部位達到了98.9％的平均精確度，整體F1-score為97.4％；DCM在分類瑕疵之身體部位達到了96.9％的準確率。本研究所提出之方法期望能夠提高台灣有色肉雞屠宰產線在瑕疵檢測上之效率與準確性。

Chicken is among the most commonly consumed meats worldwide. To provide high-quality chicken meat to customers, it is necessary to identify carcasses with defections and excluded the defected carcasses in the slaughter process. Taiwan native chickens (TNCs) are popular chicken varieties in the domestic market. Different from the broiler, TNCs include breeds with various sizes and external appearances. These varieties are usually processed in the same slaughter line. Due to these differences, TNC carcasses with defects are usually picked out in the slaughter line manually. However, manual methods are time-consuming and inefficient. Moreover, commercially available automatic carcass defect detection systems for broilers are suboptimal for TNCs due to the high variation. Thus, this study developed a machine vision system to automatically detect the TNC carcasses with defects in slaughter lines. The system comprised an image acquisition module, body parts segmentation model (BSM), and a defect classification model (DCM). The image acquisition module was water-tight and was designed to acquire complete two-sided images of TNC carcasses in a slaughter process line. The BSM was used to detect four body parts of chicken carcasses, namely the wings, breast, back, and legs. Subsequently, the DCM was used to determine whether the body parts were defected. The trained BSM achieved a mean average precision of 98.9%, and an overall F1-score of 97.4% on identifying the carcass body parts. The trained DCM achieved an accuracy of 96.9% on classifying defected body parts. The developed system is expected to enhance the efficiency and effectiveness in detecting defect parts of TNCs.