Research And Implementation Of Classification And Detection Of Plant Leaf Pest Based On TPU And FPGA

As a country with large food demand,the growth and harvest of crops are related to the basic needs of more than one billion people.In addition to the natural environment,the most serious impact on plants is the problem of pests and diseases.And such problems have many kinds,strong explosive characteristics,which often cause great losses to the national economy.Therefore,real-time and effective monitoring of plant diseases and insect pests is of great significance to China’s agricultural development.But on the other hand,with the development of urbanization,the population of farming is also decreasing,and there is a lack of enough labor to monitor the plants in the field.Aiming at this problem,under the framework of the increasingly mature Caffe deep learning technology in recent years,based on neural network Tensor Processing Unit(TPU)and FPGA SOC chip,a plant diseases classification system for edge-computing enviroment is built.The contradiction between the huge consumption of computing resources by deep learning and the limitation of computing resources in edge computing environment is solved.The power consumption,accuracy,computing speed of the system are sufficient to meet the needs of edge-side or handheld mobile-side devices.Firstly,the related principles and common network model algorithms in the field of deep learning are briefly introduced,and the basic principles,architecture,characteristics and feasibility of combining the FPGA and TPU are explored.At the same time,the Caffe deep learning framework and the data set needed in this paper are briefly introduced.Then the appropriate network model is selected and the best scheme to optimize and compress the model is explored,so that it can further meet the requirements of edge computing environment,while minimizing the impact of accuracy.Secondly,the neural network algorithm model is constructed,optimized and trained by using the Caffe deep learning tool in the PC end based on the Linux operating system.The algorithm model and related software are downloaded to the target device through Ethernet,so that the ARM CPU in the SOC chip can perform the deep neural network computing task by calling TPU relying on the API function.This paper carries TPU soft core on the FPGA SOC chip and works with the built-in ARM CPU of the new edge computing architecture to realize the acceleration calculation and accuracy improvement of the deep neural network model.And by comparing with other deep learning edge computing platforms with different computing architectures,it is proved that the system has obvious advantages in accelerating computation of small-scale deep neural networks without reducing accuracy.Finally,based on the data set of plant diseases andinsect pests,the effective diseases classification system of plant leaves are completed.The standard of accuracy,power consumption,calculation speed all meet the design requirements.