Embedded Important Target Recognition System Based On Convolutional Neural Network

Important military and civilian facilities play a very important role in morden war.With the rapid development of modern science and technology,it is particularly important to quickly and effectively control the battlefield environment and finding or fighting the goals mention above.So the role of target identification is becoming more and more important.With the rapid development of aircraft,unmanned aerial vehicles and guided weapons,On a low-power,miniaturized and low-cost hard platform,it is very practical to realize fast and accurate image targets.With the purpose mention above,this paper tries to make use of the development of convolution neural network,multicourse DSP,to make a system of important target recognition,which also can be used on the multicourse DSP made in our country,like FT-2000.In this paper,some work had been done as follow:Firstly,Making data sets for network training.On the angle of simulated aircraft,Through public channels(such as Google Earth)collected the aerial images or satellite photos about the airport,train station,dam(hydroelectric),oil deport,radio and TV Tower.According to the principle of scale similar to collect and filter these public image data.Excluding the impact of individual image classification,extended data set,through the network training test the effect of experimental,and the final data set contain that 53000 pictures in training set and 21000 pictures in the test set.Secondly,The appropriate convolutional neural network had been used to train the appropriate recognition model.Through searching literature,we can find the performance of different network framework and network model in the field of image recognition,and combined with the characteristics of using embedded devices,determine the appropriate network for model training.The transfer learning strategy is adopted to improve the training rate and efficiency,and finally an effective recognition model has obtained.Thirdly,Porting network framework and network model on embedded platform.Determining the main function of network frame function,combined with the hardware characteristics of embedded devices,without affecting the network identification function and the recognition accuracy,the source code has been cut,modify,replace,finally the target recognition function has been implemented on the embedded platform.Lastly,Optimization of target recognition function on embedded platform.The recognition rate is gradually improved by means of memory access optimization,open optimization options,multi-core parallel computing and so on.This paper is based on DSP multi-technology and convolution neural network technology,tried to solve the problem of the low speed,low accuracy and high demand of traditional image recognition technology.The convolution neural network framework is implemented in embedded devices,Perform the parallel processing performance of multi-core DSP,By optimizing the visiting memory,the adjustment optimization option achieves the improvement of target image recognition rate.