Research And Application Of Snow Surface Recognition Technology Based On Edge AI

“Edge AI”(edged AI)brings machine learning to mobile terminals,effectively declining time,economic cost and energy consumption for big data research,which is gradually becoming clear in the sight of the public.Edge AI research in quite a few industries is also gradually emerging along with the development and maturity of artificial intelligence.With the process of meterological modernization,automatic snow depth detectors have been built in many places,and the majority of areas use DSS1 type snow depth detectors independently developed and produced by the new-look aerospace science and technology Co,.LTD.,Jiangsu Provience,which play a great role in disaster prevention and mitigation by relying on the effective means of automatic snow depth measurement on frost and snowy waether,however,some erroneous data also appear from time to time.This appear focuses on the mobile terminal frost and snow recognition technology,from the perspective of machine learning,to assist the snow depth laser detector to recognize the snow on the snow on the snow measuring board and reduce the false detection rate.In this paper,an embeded Hi3559 A with deep learning module is used as the hardware processor.Aiming at the problems of small memory,low computing power,insufficient power supply and high maintenance cost of the equipment at the mobile end,a low power embeded frost and snow recognition system is designed,which mainly makes revelent improvements to snow recognition system is designed,which mainly makes relevant improvements to the image processing algorithm and hardware circuit:1.According to the principle of image compression,the original YUV data3840*2160*3 is obtained from VI,compressed into 240*240*3 and rear-ranged into YUV420 SP,which effectively reduces the amount of computation and improves the recognition rate of image classification.2.Adaptive image enhancement algorithm is designed to collect the high-frequency components in the image through high-pass filtering to make it superimposed with the original image,so that the image quality acquired by the lens in the current environment is optimal,even if it encounters frost,haze,rainfall and other weather,it still ensures the normal operation of the subsequent modules.3.Train and transplant Mobile Net V2,which is belongs to Caffe1.0 structure,to Hi3559 A camera module and to recognize the snow.4.Add a frequency dividing circuit in the hardware,to estimate the quantity of square ware from snow depth detectors,forbidding useless energy loss as well.In the test session,two main aspects were tested,including function test and performance test.The test results show that this module can realize the functions of image acquisition,image processing,image classification and GPIO output.The performance test shows that the average power is 2W,the leakage current is 0.031 A,the stand by picture 0.8s,which can be applied to the field of weather monitoring.