Research On Remote Sensing Image Aircraft Detection Method Based On Deep Convolutional Network

For the growing number of remote sensing image data,how to intelligently process,improve the efficiency of information acquisition and reduce the cost of data storage and transmission is an urgent problem.Aircraft target is a kind of information with important value in remote sensing image.This thesis mainly studies aircraft target detection in remote sensing image based on deep convolution neural network,which can carry out real-time detection on various specifications of space-based and airborne platforms.At this stage,in the field of remote sensing image aircraft target detection based on deep convolutional neural networks,lightweight model algorithms that can balance detection accuracy and speed and are suitable for different load platforms are relatively scarce.This thesis first studies the common problems in the design of deep convolution neural network,and then proposes a new backbone network and an end-to-end multi-scale detection model.In addition,this thesis studies how to improve the accuracy of the existing detection algorithm without changing the original network architecture,and makes a comprehensive innovation in three different aspects of the target detection algorithm.The main work is as follows:1.In this thesis,the common network problems and network architecture in deep convolution neural network target detection are studied.This thesis studies the causes of the problems such as network degradation and model bloated,and the principles of the current solutions,and studies the corresponding solutions for different network problems,which are integrated into the design of the new convolutional neural network.This thesis studies the architecture principle of residual network and dense connection network,and compares their adaptability in aircraft target detection of remote sensing images with specific experiments.2.Aiming at the common deep convolution feature extraction backbone network at present,this thesis proposes an identically separable backbone network ISN based on deep separable convolution and identity mapping.ISN has low parameter and computational cost,and can be applied to the detection of small load platform.Moreover,the simple connection architecture will not drag down the detection speed,and uses deep separable convolution to achieve lightweight detection and the feature extraction is enhanced by simplified identity mapping to ensure the detection accuracy of the model.Then,based on the idea of feature pyramid network,this thesis proposes a lightweight detection model MIS,which can enhance the multi-scale detection performance and achieve the balance of detection speed and accuracy.Finally,the model achieves the best performance in six groups of different model experiments,which can meet the real-time detection task requirements of different platforms.3.Aiming at the problem of improving the accuracy of target detection algorithm based on deep convolution neural network,three methods of improving the model accuracy based on non-network architecture factors are proposed,which are applied to the three stages of network model training,namely before training,during training and after training,and can effectively improve the model accuracy on the basis of the original network.(1)The depth convolution target detection algorithm based on border labeling is studied,and a fast adaptive selection method of a priori box based on Intersection over Union growth curve is proposed.This method can solve the problem that it is difficult and time-consuming to select the number of priority check boxes at the present stage.According to the data set,it can adaptively count the growth and change of the score of Intersection over Union under different number of prior boxes,so as to get the number of optimal priori boxes.Finally,the effectiveness of this method is verified by experimental comparison and analysis.(2)Aiming at the bottleneck problem that the model is difficult to further converge in the process of deep convolution network training,a training method LBT based on the joint transition adjustment of learning rate and batch size is proposed.This method can break through the training bottleneck and effectively deepen the model learning.Combined with the comparison before and after the experiment,the improvement of accuracy is confirmed.(3)Aiming at the processing defects of the non-maximum suppression method in the aircraft target detection of remote sensing image in the problem of multi frame labeling on a single target,a borders correction algorithm is proposed to correct the two defects respectively.Finally,the effectiveness of the algorithm is verified by the comparison of specific images.