Rsearch On Lightweight Pest Detection Method In Grain Storage Logistics Based On Knowledge Distillation

The rapid development of agricultural modernization and intelligence has greatly improved the grain output of our country.However,it is difficult to avoid the loss of grain due to stored-grain pest problems in grain storage.The stored-grain pest detection network based on deep learning can effectively improve the efficiency and accuracy of pest detection and identification in grain storage.At the same time,in order to improve the response speed to pest situation,the stored-grain pest detection based on mobile terminal is also of great significance.However,high performance stored-grain pest detection networks often have complex network structure and huge number of parameters,which leads to the limitation of computing and memory resources in the mobile stored-grain pest detection scenario,these networks are difficult to deploy and operate practically.However,the lightweight stored-grain pest detection network is often difficult to meet the performance requirements of practical tasks due to its low performance.In order to realize the high-performance detection and identification of the lightweight stored-grain pest detection network on the mobile terminal,a cross-layer multi-scale knowledge distillation framework is proposed in this paper,which greatly improves the performance of the lightweight stored-grain pest detection network without increasing the computational complexity and the number of parameters of the model,so as to meet the practical task needs.Therefore,the work of this study has theoretical research and practical application value.The Imain works are as follows:1.We collected and manually labeled images of stored grain pests with six different classes in the real grain storage environment,and two different stored-grain pest detection datasets were constructed.2.We proposed a cross-layer multi-scale knowledge distillation framework.The framework constructs both local structured semantic knowledge and global context knowledge,and the corresponding distillation loss functions are designed to transfer local and global knowledge to the student network to improve the performance of the student network.3.We designed a cross-layer local graph knowledge distillation module to represent and transfer local structured semantic knowledge,so as to improve the representation ability of lightweight network for important knowledge and improve its network performance.4.Furthermore,considering the overall context correlation and multiscale information of the image,this study designed a global distillation module.The global distillation module uses the dual attention graph distillation to construct the overall context knowledge of the image,and uses the multi-scale Logits distillation to construct the multi-scale output distribution knowledge,which is conducive to improving the accuracy of the lightweight network for the location and category prediction of pests.5.We conducted various experiments with our proposed cross-layer multi-scale knowledge distillation framework,which verified its ability to improve the performance of lightweight networks and its generalization ability for different lightweight networks.At the same time,the lightweight network trained based on the knowledge distillation method designed in this study was carried out a simple simulation of the storage-grain pest detection task under the mobile terminal scenario to prove the practical value of the proposed algorithm.The cross-layer multi-scale knowledge distillation framework proposed in this study can effectively improve the performance of lightweight stored-grain pest detection network without increasing the complexity and computation of the model.Experiments on two different data sets of stored-grain pest detection were carried out to verify and analyze the effectiveness of the knowledge distillation method proposed in this study,which is superior to the existing frontier knowledge distillation method for object detection task.