Beef quality dual-label classification incorporating texture and multihead map attention mechanisms

Abstract

With the increasing consumer demand for high-quality and safe beef, there is an urgent need for advanced scientific testing methods to ensure product quality. This study introduces a rapid and accurate dual-label classification system for beef cuts and marbling grades, combining a multi-head attention mechanism and contrast features extracted from the Gray-Level Co-occurrence Matrix (GLCM), significantly improving classification accuracy. A comparison of different ResNet models revealed that ResNet50 achieved the highest classification accuracies, 85.65 % and 85.03 %, respectively. Building on the ResNet50 model, GLCM was used to extract contrast features in four directions from beef images, followed by feature fusion. Compared to SE and non-corresponding multi-head Graph Attention Networks (GAT), a multi-head GAT focusing on GLCM texture features in each direction was selected, with these features fused one-to-one with the fully connected layer features of RGB images and incorporating a feature fusion mechanism. Our model achieved 93.5 % accuracy in cut classification and 92.25 % in grade classification. Additionally, an app based on the optimal model was developed, enabling users to perform real-time testing and obtain results. The goal of this study was to develop the fastest and most accurate dual-label classification system, significantly improving the speed and reliability of obtaining product information during marbling selection.