Bai Zongxiu, Zhu Rongguang, He Dongyu, Wang Shichang, Huang Zhongtao
College of Mechanical and Electrical Engineering, Shihezi University, Shihezi 832003, China.
Key Laboratory of Northwest Agricultural Equipment, Ministry of Agriculture and Rural Affairs, Shihezi University, Shihezi 832003, China.
Foods. 2023 Sep 27;12(19):3594. doi: 10.3390/foods12193594.
To achieve accurate detection the content of multiple parts pork adulterated in mutton under the effect of mutton flavor essence and colorant by RGB images, the improved CBAM-Invert-ResNet50 network based on the attention mechanism and the inversion residual was used to detect the content of pork from the back, front leg, and hind leg in adulterated mutton. The deep features of different parts extracted by the CBAM-Invert-ResNet50 were fused by feature, stitched, and combined with transfer learning, and the content of pork from mixed parts in adulterated mutton was detected. The results showed that the of the CBAM-Invert-ResNet50 for the back, front leg, and hind leg datasets were 0.9373, 0.8876, and 0.9055, respectively, and the RMSE values were 0.0268 g·g, 0.0378 g·g, and 0.0316 g·g, respectively. The and RMSE of the mixed dataset were 0.9264 and 0.0290 g·g, respectively. When the features of different parts were fused, the and RMSE of the CBAM-Invert-ResNet50 for the mixed dataset were 0.9589 and 0.0220 g·g, respectively. Compared with the model built before feature fusion, the of the mixed dataset increased by 0.0325, and the RMSE decreased by 0.0070 g·g. The above results indicated that the CBAM-Invert-ResNet50 model could effectively detect the content of pork from different parts in adulterated mutton as additives. Feature fusion combined with transfer learning can effectively improve the detection accuracy for the content of mixed parts of pork in adulterated mutton. The results of this study can provide technical support and a basis for maintaining the mutton market order and protecting mutton food safety supervision.
为了通过RGB图像准确检测在羊肉香精和色素作用下掺假羊肉中多个部位猪肉的含量，使用基于注意力机制和倒置残差的改进型CBAM-Invert-ResNet50网络来检测掺假羊肉中背部、前腿和后腿的猪肉含量。通过特征融合、拼接并结合迁移学习，融合CBAM-Invert-ResNet50提取的不同部位的深度特征，检测掺假羊肉中混合部位的猪肉含量。结果表明，CBAM-Invert-ResNet50对背部、前腿和后腿数据集的准确率分别为0.9373、0.8876和0.9055，均方根误差值分别为0.0268 g·g、0.0378 g·g和0.0316 g·g。混合数据集的准确率和均方根误差分别为0.9264和0.0290 g·g。当融合不同部位的特征时，CBAM-Invert-ResNet50对混合数据集的准确率和均方根误差分别为0.9589和0.0220 g·g。与特征融合前构建的模型相比，混合数据集的准确率提高了0.0325，均方根误差降低了0.0070 g·g。上述结果表明，CBAM-Invert-ResNet50模型能够有效检测掺假羊肉中作为添加剂的不同部位猪肉的含量。特征融合结合迁移学习可以有效提高掺假羊肉中猪肉混合部位含量的检测准确率。本研究结果可为维护羊肉市场秩序和保障羊肉食品安全监管提供技术支持和依据。
