Remote Sensing Image Object Detection And Recognition Based On Convolutional Neural Network And Data Augmentation

Remote sensing images have the characteristics of large image size and small proportion of object relative image.Whether it is a SAR image or an optical remote sensing image,the features contained in it have extremely high research value in military and urban management.With the development of remote sensing satellite technology,the resolution of remote sensing images continues to increase.However,it is difficult to obtain labeled SAR image data sets for object detection and recognition.This thesis aims to expand the SAR image and optical remote sensing image data sets using data augmentation based on convolutional neural network,and proposes improved methods for existing problems.The main work of this thesis is divided into three parts.(1)Aiming at the problem of limited amount of data for SAR image object detection and recognition,a new data set expansion method based on convolutional neural network is proposed.First,based on the SAR image military objects and scene pictures in the MSTAR data set,the SAR image object detection and recognition data set SAR_OD was produced.Input the image in the SAR_OD data set into the convolutional neural network designed in this thesis,determine the position of the picture suitable for placing the new object,and place a suitable number of objects on the picture to obtain the data set SAR_OD+.The experiment proves that the training model of the data set SAR_OD+ has a significant improvement in the evaluation index,especially in the experiment of 50% training data,the improvement effect is more significant.Therefore,in the case of limited data,the data set expansion method can be used to further improve the performance of SAR image object detection and recognition.(2)In order to solve the problem of poor classification effect of some class samples,this thesis uses an improved cross-entropy loss function to replace the original classification loss function of Faster R-CNN,to further improve the contribution of difficult to classify samples to the loss function,so as to achieve better object detection and recognition results.Experiments show that this method has improved in different degrees on the models trained in each data set,and the object category with low average accuracy has been improved significantly.The improved Faster R-CNN is trained through the SAR_OD+ dataset,and the trained model is used to build a SAR image object detection and recognition system based on Py Qt 5,which implements the model performance testing function and single image detection function.(3)Aiming at the characteristics that the background of optical remote sensing images is complex and the background of each type of object is relatively uniform,an optical remote sensing image data augmentation method based on convolutional neural network is proposed.First,select the images of aircraft,ships and vehicles in the optical remote sensing image data set NWPU-VHR 10 to make the data set VHR-3.Input the image to the convolutional neural network model to predict the object category that should be placed,and input the image to different networks according to the results,determine the suitable location to place the object and place the object of a specific category.According to the characteristics that some objects in the data set are too dense and there are many overlapping parts of the labeling frame,an improved non-maximum suppression method is used to remove the redundant candidate frame.The experimental results show that the improved non-maximum suppression method can effectively improve the average precision of the categories where the object frame is easy to overlap.After the data augmentation,the recognition performance is further improved,and when the amount of data in the training set is less(50% of the data),the improvement effect is more significant.