Research On Aircraft Detection Method In Remote Sensing Image Based On CNN

With the development of science and technology,there has been a great breakthrough in improving the resolution of remote sensing image.The high-resolution image contains a lot of feature information.Especially in the field of remote sensing research,it is necessary to focus on analyzing the spatial structure and shallow texture characteristics of ground objects in the image,mainly to lay the foundation for extracting finer ground features.Remote sensing image object detection is one of the main categories of high-resolution remote sensing image applications.As a key object in civil life and military operations,aircraft has certain theoretical significance and practical value for its detection.In this essay,firstly,using a deep convolutional neural network to extract high-resolution image feature information,and classify scenes based on differences in features.Then,aiming at the problem that the result of aircraft detection in remote sensing image is not good at present,a method of aircraft detection based on deep convolution neural network is proposed.Finally,we conducted experimental tests to complete the aircraft detection task in the airport scenario,and achieved good results.The main work content and results are as follows:The first part is about classification of remote sensing scenes.Making use of the good classification ability of convolution neural network and the advantage of transfer learning,this essay proposes a PISC method based on Inception-v4 network to identify various scenes in high-resolution remote sensing images and obtain airport scenes.Firstly,the pre-training model and the corresponding initialization parameters are obtained by training Inception-v4 of the deep convolution neural network on Image Net,which can greatly reduce the consumption of training time and avoid the low classification accuracy caused by overfitting phenomenon.Then the remote sensing data of small samples were divided into the pre-training model in proportion,and the network parameters were constantly adjusted according to the characteristics of the sample set to obtain the best classification model.Finally,compared with the existing scene classification methods,the experiments show that the method of this essay achieves an accuracy of 97.92% on the UC Merced Land Use scene image data set,which effectively improves the high-resolution image scene classification accuracy.The second part is about aircraft detection in airport scenarios.In the case of obtaining the target area of the airport scene,the object detection model is introduced to detect the aircraft in the airport scene.Aiming at the problem of poor evaluation indicators such as F1-score and Recall rate during the detection of Yolov3,we put forward an improved Yolov3 remote sensing image aircraft detection method.In this essay,the Inception module is introduced into the Yolov3 network structure to enhance the precision and m AP values,and improve the object detection effect.Then,by using the K-means algorithm to cluster the data set,the network parameters are adjusted and the input image resolution is improved in the pre-training model.The best aircraft detection model is obtained by using multi-scale training method.The training experiments on the remote sensing aircraft dataset in RSOD-Dataset show that the Inc-Yolov3 algorithm proposed in this essay has improved the accuracy,Recall,F1-score,IOU and m AP in the remote sensing aircraft detection.This method can be applied to detect other ground objects,which has certain practicability and popularization value.Get the airport scene from the image scene classification,and then detect the aircraft target from the airport scene.It accomplishes a complete aircraft object detection work,which is an important step for the aircraft object detection algorithm to enter the practical application.