Identification and Analysis of Pork Freshness Quality Based on Improved MobileNetV3

Highlights Model can perform pork freshness quality testing under complex lighting. MobileNetV3_E outperforms other state-of-the-art models. Model achieve an optimal balance between performance and volume. Model supports deployment on edge devices. Abstract. The accurate appraisal of freshness is influenced by the color of the meat, which is a critical indicator of pork’s freshness. However, lighting changes can also influence consumers’ perceptions of meat color. To address this issue, this study recommends a pork freshness detection method based on the lightweight MobileNetV3 model that can effectively handle interference from complex lighting environments. In this study, a lighting factor is introduced into the original dataset. The results are compared with experiments without lighting interference, showing that lighting interference leads to varying degrees of model performance degradation. This investigation introduces Efficient Channel Attention (ECA) into the MobileNetV3 model (MobileNetV3_E), redefines the h-sigmoid and h-swish activation functions, and fine-tunes the model training using the official pre-training weights to achieve the optimal balance between model performance and volume. Only the ECA module-related layers and custom fully connected layers are trained. The experimental results show that MobileNetV3_E is less affected by the lighting interference and still maintains exceptional performance on the test set, where the accuracy is 98.6%, the loss function value is 0.069, the precision is 98.6%, the recall is 98.5%, and the F1-Score is 98.5%. The MobileNetV3_E model achieves a compact size of 17.34 MB, making it suitable for deployment on resource-constrained devices. It indicates that the MobileNetV3_E model provides an efficient, reliable and easy-to-deploy solution for pork freshness detection under complex lighting conditions, which is of great application value. Keywords: Deep learning, MobileNetV3 network, Non-destructive testing, Pork freshness.