Improved Deep Convolution Neural Network And Its Application In UAV Target Recognition

In recent years,deep learning in UAV perception and avoidance is a hot research direction.The deep learning method is used to analyze the images captured in the UAV vision range,which can better perceive the surrounding environment,effectively avoid obstacles and plan better route.This paper studies and analyses the development and application of deep learning in image classification,target detection,face recognition and key point detection.The existing deep learning algorithm is improved and applied to UAV target recognition,and good results are achieved.The main contents of this paper are as follows:(1)A multi-strapdown residual network model is proposed and the classical deep residual network is improved on the basis of deep convolution neural network(DCNN).The network models with different depths and parameters are studied.A multistrapdown residual network model(Mu-ResNet)is designed by combining two-layer residual learning module with three-layer residual learning module.The experimental research was carried out by constructing multiple strapdown residual networks with different depths,the final test results show that the overall performance of Mu-ResNet is better than that of ResNet.(2)The classic Regional Proposal Network(RPN)was improved and a search mode for the Cascading Regional Proposal Network(CRPN)was proposed.The combined Cascading Regional Proposal Network(CRPN)is combined with the MultiStriped Residual Network(Mu-ResNet).Since the cascading area suggests that the network searches for the feature maps of different sizes,the calculation amount is greatly reduced,and the effect of detecting small targets in the image is better.The network model combined with Mu-ResNet and CRPN is used for multi-target recognition test on the UAV target data set and PASCAL VOC data set.Compared with the network model using ResNet and RPN,the recognition accuracy is greatly improved.(3)Aiming at the characteristics of face recognition and face key detection,based on multi-strapped residual network and cascaded area suggestion network,a face recognition algorithm based on feature point correction is designed to identify target classification.The pedestrians detected in the classification are finely classified(face recognition).The algorithm performs face recognition and face key point detection with the same convolutional neural network,and corrects the face recognition frame by detecting the key position.Experiments show that our algorithm can get better on the FDDB data set.Identify accuracy,recall,and average error at key points.The face recognition of pedestrians in the target data set of the drone has also achieved good results.All the algorithms proposed in this paper are developed and debugged by using MATLAB,Python and C/C++ language on a GPU with 8g memory.UAV target detection data sets are a large number of UAV aerial images.The algorithm not only completes simulation experiments on its own data sets,but also validates them on public data sets(cifar10,PASCAL VOC,FDDB,etc.).The high accuracy and real-time performance of the algorithm in data sets are two important metrics concerned in this paper.The real-time performance and accuracy of the algorithm are tested in the experimental part of each chapter.Finally,the test on a large number of data sets achieves the expected experimental results.