Grade-Mamba: Efficient Visual Representation Learning for Beef Carcass Classification of Brown Cattle

As global beef production continues to grow, the efficiency and accuracy of carcass assessment and grading have become critical within the broader agricultural and food economy framework. The price and classification of the meat derived from the inspected carcass will depend on this inspection. Currently, carcass assessment primarily employs two detection methods: convolutional neural networks (CNNs) and vision Transformers. Unfortunately, CNNs have limited ability to capture global contextual relationships due to their local receptive fields, which negatively impacts classification accuracy. Meanwhile, vision Transformers face scalability issues due to the computational demands of their attention modules, which grow in O(n²) complexity, posing significant challenges for resource-constrained practical applications. To address the inefficiency and subjectivity of manual inspection in large-scale meat grading, while adapting to hardware environments with limited GPU memory and computing power, we propose Grade-Mamba, an automated vision system for grading brown cattle carcasses. Grade-Mamba achieves both high accuracy and computational efficiency through two key innovations. First, it employs an improved squeeze-excitation (SE) module to dynamically enhance key texture and color features (such as marbling) and improve feature relevance. Second, it utilizes dynamic spatial pyramid pooling (DSPP) to capture multi-scale spatial hierarchies for robust feature extraction of diverse carcass characteristics. By integrating these components, Grade-Mamba provides real-time, accurate grading while minimizing resource requirements, becoming a scalable and objective alternative to traditional methods. This integration enables the system to perform high-throughput meat quality assessment in resource-constrained environments, significantly boosting production efficiency while maintaining precise grading performance. Grade-Mamba enables sub-4-second automated grading per carcass, offering a scalable, objective, and high-throughput solution for modern meat processing facilities.