A multi-task PSO-Attention framework for joint marbling segmentation, object detection, and species classification of meat [pdf]

Background.Accurate assessment of meat quality requires complementary visual information from marbling segmentation, anatomical part detection, and species classification.However, existing approaches treat these as independent tasks, leading to inefficient and fragmented workflows.Current deep learning models also suffer from extensive manual hyperparameter tuning and lack a unified framework that simultaneously addresses all three tasks.Material and methods.This study proposes a novel multi-task PSO-Attention framework that jointly performs marbling segmentation, anatomical part detection, and species classification using a shared Effi-cientNet-B0 backbone enhanced by Convolutional Block Attention Modules (CBAM) and Particle Swarm Optimization (PSO) for automated hyperparameter search.The framework was validated on three independent meat-image benchmarks covering pork, beef, and lamb for segmentation, chicken parts for detection, and pork/beef/horse for classification.Results.The proposed framework demonstrated consistent performance improvements across all three tasks.For marbling segmentation, the UNet+PSO-Attention configuration achieved the highest mIoU of 0.9362, outperforming the baseline UNet (0.8940) by 4.22 percentage points.For anatomical part detection, our method attained the best mAP@75 of 0.7904, surpassing the strongest baseline YOLOv8n (0.7775).For species classification, the framework achieved an accuracy of 0.9671, comparable to the single-task ResNet-50 baseline (0.9726) while requiring only one unified training pipeline.Ablation studies confirmed the synergistic contribution of CBAM and PSO modules across all tasks.Conclusion.The proposed PSO-Attention framework serves as an effective, plug-and-play multi-task solution that simultaneously enhances feature discrimination and optimizes training parameters, replacing three independent models with a single deployable system.It provides a robust and scalable approach for comprehensive meat quality assessment, demonstrating how combining swarm intelligence with attention mechanisms can consistently elevate deep learning performance in food image analysis.