Non-destructive detection of pork loin freshness based on the YOLO v11 model

ObjectiveTo develop an intelligent method for non-destructive recognition of pork loin freshness based on computer vision technology, which is to be achieved by establishing a classification model for accurate, fast, and non-destructive detection of pork freshness in the cold chain industry.MethodsA classification detection model is proposed based on the YOLO v11 architecture. First, total volatile basic nitrogen (TVB-N) content and pH values are measured before the meat samples are divided into three categories: "fresh", "less fresh", and "spoiled". A dataset of 4 680 high-resolution images is constructed (generated from 1 247 original images enhanced through operations such as rotation, flipping, cropping, brightness adjustment, and blurring). Then, the dataset is divided into training, validation, and test sets in a 6∶2∶2 ratio. Utilizing a transfer learning strategy, the YOLO v11 model with pre-trained weights is initialized in the PyTorch framework and trained using the SGD optimizer (learning rate of 0.01). Finally, the model established is compared with six mainstream models: DenseNet161, EfficientNet_b2, MobileNet_v2, ResNet50, ShuffleNet_v2_x1, and ViT_base_patch16_224.ResultsThe YOLO v11 model achieves excellent performance, with an overall classification precision of 1.000, significantly outperforming the best baseline model, ViT_base_patch16_224. Additionally, the proposed model demonstrates significant advantages over other models in both F1 score and accuracy metrics.ConclusionIn this study, the YOLO v11 model is successfully applied to the non-destructive detection of pork loin freshness. With high-precision generalization capabilities, the proposed instance classification method achieves a detection rate of 1.000 for spoiled meat and ensures food safety. With an inference speed of 12 ms, this model is suitable for real-time non-destructive detection in production lines. Additionally, exhibiting a strong generalization ability, the model can evenly recognize multiple levels of freshness, with classification accuracy greater than 98% for each category.