Research On Convolutional Neural Network Algorithm And Its Application

The convolutional neural network(CNN)is a new stage in the development of artificial neural network(ANN)and also a product of large data era.In the light of biometric information,the network adopts local connectivity,weight sharing and sub-sampling to reduce the complexity of the convolutional neural network.At the same time,the CNN network structure has the displacement invariance,deformation invariance and the scale non-deformation.At present,convolutional neural network has made historic achievements in many areas,including natural language processing,computer vision and biometric identification.With the drastic increase of computer performance,the new convolutional neural network becomes deeper to handle the massive data.Compared with millions of contest datasets,small datasets in industrial are difficult to adapt to the existing deep convolutional neural networks.Based on the comprehensive understanding of convolutional neural networks,this dissertation studies the related theories thoroughly.Combined with the transfer learning theory and the development of the convolutional neural networks in recent years,the classification and the aircraft target detection of remote sensing images are studied.The research work can be summarized as follows:Firstly,aimed at the problem that small dataset cannot be directly trained in the complex depth convolutional neural network,the classification of remote sensing images is studied by pre-trained VGG-16,GoogleNet,and ResNet-152 models.Through simulation test,the pre-trained ResNet-152 achieves the best classification accuracy of 98.7% and 99.3% in the remote sensing image dataset UCMerced and SIRI-WHU datasets.The accuracy of classification is the highest classification accuracy achieved on the datasets and it further verifies that transfer learning can make deep convolutional neural networks well adapted to small datasets.Secondly,the convolutional neural network is applied to solve the difficult problem of target detection,aircraft target detection in remote sensing images.According to the characteristics of the aircraft to be detected,the parameters of the existing target detection network Faster R-CNN are modified to adapt to the detection tasks of large range and small target in remote sensing images.In order to meet the requirement of the detection model for the data amount,1000 aircraft remote sensing images of different resolution are collected as experimental dataset from the Google Earth.Compared with the original Faster R-CNN and the existing excellent detection network model,the simulation results on the collected dataset show that the modified network has a better recognition effect for the aircrafts in the remote sensing images with a lower false alarm rate and the requirement for real-time can be met simultaneously.