Typical Military Target Recognition Based On Image Processing And Deep Learning

Under modern warfare,satellite radars,unmanned aerial vehicle reconnaissance platforms,image precision guided munitions,etc,have been developed and applied.With the rise of information warfare,the battlefield environment is becoming increasingly complex and changeable,and the form of combat is undergoing profound changes.Use a camera or other imaging equipment to capture images,effectively process the acquired images,and calculate the type,position,speed and other information of the expected target from them.Making full use of these accurate information is of great value for command assistance decision-making and precise strikes..Based on deep learning and image processing theory,this paper has carried out research on automatic recognition technology of military targets,and completed the classification and detection tasks of typical military targets such as tanks,aircraft,artillery,armored vehicles,ships,ports,buildings,and soldiers.First,based on the research on the theory of convolutional neural networks,this paper proposes an improved activation function Re LU-Xe X and an improved combined convolution module.The simulation experiment demonstrates that the improved activation function Re LU-Xe X improves the accuracy of the model more significantly than other common activation functions;the simulation experiment demonstrates that the improved combined convolution module is compared to VGGNet（standard convolution）,Goog Le Net（Inception module）,Mobilenet（depth separable convolution）,Res Net（residual error module）,etc,have higher accuracy,and the speed is between standard convolution and depth separable convolution.Secondly,researches on visible light image dehazing,SAR image despeckle and military target image classification are carried out.Based on the improved combined convolution module,a military target classification network is designed,and classification simulation experiments are completed on three data sets containing typical military targets: MSTAR,self-made,and UCMerced.In the three data sets,compared with other methods,the method in this paper has achieved the highest accuracy and better speed.The test accuracy rates under the three data sets have reached99.753%,91.718%,and 92.381%,respectively.Compared with standard convolution,the speed is increased by about 20%.Combining the improved activation function Re LU-Xe X with the improved combined convolution module,a fully convolutional neural network is designed,and the accuracy of the combined convolution module and Re LU-Xe X activation function is further verified on the UCMerced data set.But at the same time,it is also concluded that the Re LU-Xe X activation function will have a certain adverse effect on the model speed.Based on the prior defogging of dark channel,the comparison and analysis of classification accuracy before and after defogging are completed.Finally,based on the DOTA high-resolution aerial scene target detection dataset,the YOLOv5 model was trained.After 100 rounds of training,the accuracy of the model was 54.76%,the recall rate was 81.47%,and m AP@0.5 reached 74.12%.The detection and analysis of three potential military target scenes of ship ports,airports and seaports have been carried out.The detection results can still be achieved in high-resolution aerial scenes,and the average test speed has reached 181.8FPS.Subsequently,an aerial photography military target detection system was built based on Jetson TX2 and UAV,and the model reasoning process was completed on Jetson TX2.After compressing the image input size,the detection speed on the embedded platform reached 16.13 FPS.