Research On Pork Primal Cuts Determination Method Based On Convolutional Neural Network

Pork is the most consumed meat in the world and dominates meat consumption in China.With the development of the economy and the continuous improvement of people’s living standards in China,people are paying more attention to the quality of pork.There are many factors that affect the quality of pork primal cuts,and the cutting position is one of them.The quality and taste of meat from different cuts of the same pig’s body also vary and are suitable for different cooking methods.Traditional methods for classifying pork primal cuts require professionals to use equipment to measure and analyze the elemental composition to achieve the purpose of pork primal cuts classification.The experimental steps are cumbersome,and the real-time performance is not high.In recent years,with the rapid development of deep learning and the continuous improvement of hardware device computing power,deep learning has achieved tremendous success in many different fields.Convolutional neural networks are one of the most successful methods of deep learning in the field of image vision and have been widely used in image classification problems.This study constructed a image dataset of pork primal cuts for convolutional neural networks and designed a network model that can automatically classify and identify pork primal cuts using the principles of convolutional neural networks in deep learning theory.To address the gap of existing datasets in pork primal cuts classification,this study collected images of pork primal cuts,expanded the sample size through data augmentation,and constructed a image dataset of pork primal cuts.A convolutional neural network model combining knowledge distillation and the attention mechanism CBAM was designed and applied to train and recognize pork primal cuts classification images.The specific work contents are as follows:(1)The use of image classification methods in deep learning for experiments was proposed.Various image classification algorithms were trained and recognized on pork data set images,and their recognition accuracy was compared to select the best-performing network as the baseline model.The experimental results showed that the ResNet series models performed the best in pork primal cuts recognition,and ResNet-101 had the highest recognition accuracy,but its model network was more complicated.(2)The use of knowledge distillation to compress the model was proposed.The best-performing ResNet-101 model was selected as the teacher network,and the simpler ResNet-18 and GoogleNet were selected as the student networks for knowledge distillation.The experimental results showed that the student network model’s performance was improved after knowledge distillation,indicating that knowledge distillation can compress the model with little loss of accuracy and is suitable for pork primal cuts classification research.(3)A method combining knowledge distillation and the attention mechanism CBAM was proposed to improve the model’s recognition accuracy and reduce the model’s computational complexity.The CBAM-ResNet50,which is fused with the attention mechanism CBAM,was used as the teacher network.Compared with the previous knowledge distillation results,it was found that the distilled effect of the student network ResNet-18 with the teacher network CBAM-ResNet50 was the best,and the student network’s recognition accuracy reached 93.98%.Finally,a recognition model with balanced classification performance and model complexity was obtained.To address the problem of difficult to distinguish pork primal cuts,this study proposed a computer vision-based method to identify different images of pork primal cuts.Four different pork primal cuts(ham,belly,loin,and neck)were used as experimental data,the results show that the method proposed in this paper can identify pork primal cuts well,which has certain research significance and practical value.It also demonstrates the application potential of computer vision technology in various fields.