Ultra-Fast AR Grading System: Real-Time Beef Quality Grading With Micro Attention U-Net and VGG-16

Speed of beef grading is crucial. Grade determination must be completed quickly, because all daily quantities must be graded within a set time each day. However, current grading system is slow and inconsistent, with grading times ranging from 2 to 5 minutes per beef carcass. Current system is passive, in which beef graders analyze the beef manually. Therefore, we propose Ultra-Fast AR Grading System (UFAR Grading System), an automated system for beef quality grading. The system consists of two parts: Segmentation and Classification. In the Segmentation part, the segmentation model operates when the beef grader captures a beef image with an AR device. The segmentation model segments the muscle area, which is crucial for determining the beef quality grades. In the Classification part, once this muscle area image is sent to the server, the classification model is activated to grade beef quality. To implement Ultra-Fast AR Grading System, a fast and lightweight segmentation model that is suitable for AR devices was required. Therefore, we developed Micro Attention U-Net (MAU-Net) model based on Attention U-Net architecture. MAU-Net achieved 0.25M parameters and 5.88ms inference speed. Furthermore, we aimed to find a classification model that could quickly and accurately classify grades using the muscle area images obtained from MAU-Net. We experimented with three classification models, ResNeXt-50, DenseNet-121, and VGG-16. Among these, VGG-16 exhibited the fastest inference speed of 2.59ms. And in combination with MAU-Net, it achieved a grading accuracy of 98.283%.