Research On Lead-Zinc Ore Image Classification Method Based On Knowledge Distillation Theory

Lead-zinc ore is a strategic resource,widely used in various industrial fields,and plays a decisive role in the rapid development of the national economy.With the depletion of high-grade mineral deposits,the utilization rate of lead-zinc ore needs to be improved,and the sorting of lead-zinc ore has attracted more and more attention.Through lead-zinc ore sorting,enrich lead-zinc ore and improve utilization rate,enterprises can achieve green development without waste.In order to solve the problem to improve the accuracy of ore classification and process capacity at the same time,when using convolutional neural networks to intelligently sort ores,a lead-zinc ore image classification method based on knowledge distillation theory is proposed.In this paper,based on X-ray imaging technology of lead-zinc ore images as the research object.By studying the traditional convolutional neural network model,the knowledge distillation theory is introduced to transfer the model to achieve the effect of model optimization.Finally,extract the image feature information,use the optimized model classifies lead-zinc ore images,reduce the system resource usage and improve classification accuracy.The main contents of this paper are as follows:(1)The characteristics of image data of lead-zinc ore based on X-ray imaging technology are summarized.The properties of lead-zinc ore and the basic principle of X-ray imaging are introduced.Finally,the gray value and size distribution of lead-zinc ore images are counted to obtain the image data characteristics.(2)Construction of simulation experiment platform.Based on Caffe framework,a simulation experimental platform for lead-zinc ore separation was built.GPU was used to accelerate the recognition speed of the model and realize real-time image classification of lead-zinc ore.(3)The classification experiment of lead and zinc ore image with traditional convolutional neural network model is summarized.Introduce the structure and basic principles of the traditional convolutional network model in detail.According to the comparison of the experimental results of the traditional convolutional neural network model on the lead-zinc ore image classification,it is concluded that the recognition speed of the LeNet-5 network model is the fastest,and the recognition effect of the ResNet18 network model is the best.So,design a model that combines LeNet-5 and ResNet18 network models can meet the research purpose.(4)Based on the theory of knowledge distillation,establish a LeNet-ResNet18 network model.Change the activation function in the traditional LeNet-5 network model,use the teacher model ResNet18 to "teach" the student model LeNet-5 to learn,improve the classification performance of the network model,and solve the problem of calculation more by distillation,make LeNet-ResNet18 network model have a high efficiency and real-time at the same time.Eventually achieves the research purpose of simultaneously improving the processing capacity and classification accuracy of lead-zinc ore.