ResNet and Yolov5-enabled non-invasive meat identification for high-accuracy box label verification

Abstract

Compliance issues riddle the agricultural sector despite being an essential industry for the human race. Many factors contribute to compliance issues; however, meat cut label verification is one of the most critical concerns due to its erroneous nature. In addition, meat cut identification is complex for the human eye. Thus, access to a skilled labor force is challenging. Nevertheless, meat compliance is essential since it has export compliance ramifications. These factors, along with others, are pushing for a digital alternative. An alternative that can augment human decision-making in verifying meat cut box labels. Artificial Intelligence (AI) is a digital alternative that can boost quality assurance tasks. One of AI’s potential quality assurance solutions is a Meat Box Labeling Verification (BLV) solution that can verify the boxed meat type against the label to ensure no mismatch. Two major components make up the BLV solution: Meat Identification and Label Analysis. This work aims to solve the former component by exploring different meat-type identification techniques. It explores them by building, evaluating, and testing different computer vision solutions. Hence, the novel contribution of this work is three folds. We first build, test, and design computer vision solutions that detect meat boxes and pieces accurately. These solutions include deep learning methods in classification and object detection techniques. Following that, we evaluate these models and derive key insights. Such insights are valuable for prototyping such solutions in production environments. For example, classification models achieve a 99% testing accuracy in identifying box types. In contrast, object detection achieved 89% mAP@.5:.95 in identified individual meat cuts. Finally, we prototype an object detection model in a natural meat processing environment—the demonstration showed comparable object detection precision at 85% mAP@.5:.95.