An Integrated Approach to Enhancing YOLOv11 for Red Meat Classification Using Ghost Convolution and Optimized Conv, C3k2, and C2PSA Blocks

Abstract

This study investigates Ghost-based optimization of YOLOv11 for fine-grained red-meat classification involving beef, lamb, pork, and background classes under resource-constrained conditions. Two proposed configurations are examined, namely GhostConv and Full Hybrid, where the latter integrates GhostConv with C3k2Ghost and C2PSAGhost to jointly enhance redundancy reduction, multi-scale feature representation, and attention-based feature selectivity within a unified architectural framework. Unlike prior lightweight adaptations that apply efficiency modules in isolation, this work systematically evaluates both configurations across all YOLOv11 scales (n, s, m, l, and x) through ablation experiments, repeated runs with multiple random seeds, and statistical robustness analysis. The results indicate that GhostConv consistently delivers strong classification performance across several scales, achieving the highest mean F1-scores on YOLOv11-n, YOLOv11-s, and YOLOv11-l, whereas Full Hybrid offers the most favorable balance between predictive performance and computational efficiency at larger scales. In particular, at YOLOv11-x, Full Hybrid attains a mean accuracy of 96.47% and a mean F1-score of 96.58% while reducing model parameters from 51.603M to 14.128M and GFLOPs from 14.154 to 5.497 relative to the baseline. Further analyses based on effect size, training curves, confusion matrices, and qualitative outputs confirm that both GhostConv and Full Hybrid improve convergence stability, class-level discrimination, and representational robustness. These findings demonstrate that Ghost-based architectural optimization can improve both classification performance and computational efficiency, providing an effective solution for practical red-meat authentication and image-based food inspection systems.