Research Of Object Detection Technology Based On Convolutional Neural Networks

With the continuous development of deep learning technology in recent years,computer vision technology has also been combined with it,and has achieved leapfrog development.Object detection is a very important part of the field of computer vision,so object detection technology based on deep learning is a very valuable research field,and has achieved many research results.In remote sensing images,the effective recognition of vehicle targets is of great significance,and has application value in the field of traffic control and military reconnaissance.In this thesis,a target recognition algorithm based on deep convolution neural network is used to effectively recognize vehicle targets in aerial photographs during the day and at night,and the problem of insufficient data is solved by using migration learning.The main contributions of this thesis are as follows:(1)Collect the aerial data set of the Chinese Academy of Sciences,supplement and collect the data set from Google Earth,and label the location information of the vehicle target.Under Ubuntu 16.04 system,a Caffe deep learning framework is built,and Faster R-CNN algorithm is applied to effectively identify daytime aerial vehicle.(2)The data set of night aerial photography is collected and labeled by UAV,and the iterative algorithm based on Retinex is selected to deal with it,so as to weaken the influence of illumination and facilitate migration and learning with the data set of source domain.(3)Conditional generative adversarial network is used to preprocess night aerial images instead of Retinex iterative algorithm,which can greatly reduce processing time,speed up preprocessing and facilitate fast target detection.(4)The migration learning is applied to the night data set,which solves the problem of insufficient data and identifies the vehicle target quickly and effectively.The effectiveness of this transfer learning method is proved by comparative experiments.