Research On Detection Algorithm Of Parking Aircraft In Optical Remote Sensing Image

Parking aircraft target detection is of great significance to airport traffic control,key military base deployment,monitoring enemy real-time military dynamics and other military and civilian fields.Parking aircraft target detection and recognition based on space-based optical remote sensing image has become a research hotspot in the field of optical remote sensing image processing technology.There are still the following problems in the detection of berthed aircraft targets based on high-resolution optical remote sensing images :(1)the area proportion of airport berthed aircraft targets in high-resolution optical remote sensing images is small,which is prone to the problems of computational redundancy and the imbalance of positive and negative sample categories;(2)There are many kinds of targets,different sizes,random orientation and dense arrangement of aircraft parked at airports,which put forward higher requirements for detection algorithm;(3)The target background characteristics of space-based optical remote sensing images are complex,and there may be problems such as cloud blocking and interference from image quality degradation factors,which can easily lead to the degradation of detection performance of the algorithm.In view of the above problems,this paper focuses on airport area extraction,berthed aircraft target detection,space-based observation complex scene imaging simulation data generation and other research work.The specific work content is as follows:(1)An airport region extraction method based on the significance of straight line segments is proposed.In view of the small area proportion of the parking aircraft in the remote sensing image with large field of view,the direct detection of the full image will have a large number of redundant calculations,and the pre-detection of the airport area can effectively solve this problem.Is firstly analyzed the main features of the airport area,on this basis,design a fast line segment detection operator LSD was carried out on the runway feature extraction,the algorithm is sensitive to linear characteristics and small amount of calculation,and combined with the density of significant principle of line segment density is presented based on the significance of airport area extraction method,which can realize fast and exact extraction of airport areas.(2)A convolutional neural network based airport parking aircraft target detection method is proposed.Taking advantage of the advantage that convolutional neural network can extract deeper semantic features of targets,this paper designs a cross-scale linking feature pyramid module that can integrate different semantic levels,which can realize effective feature extraction for multi-type,multi-scale and densely arranged berthed aircraft targets.Moored on the plane,and the problem of random samples of plus or minus class imbalance problems,put forward the dynamic anchor learning module,by setting the matching degree of label distribution and matching sensitive loss function,enhance the capacity of network training,cutting rotation number of anchors,reduce the class imbalance problem,can be realized in the airport area parking aircraft target of high detection accuracy.(3)A data set construction method based on in-orbit imaging simulation of optical remote sensing is proposed.In this paper,the image interference factors in remote sensing detection link are analyzed and the image quality representation model of detection link is established.On this basis,combined with the existing open source data set,the optical remote sensing data set of the berthed aircraft under the condition of on-orbit imaging is simulated,which provides data input for the training of the berthed aircraft target detection algorithm and the verification of the algorithm performance.(4)Experimental verification and analysis.The training set was constructed from the simulation data of space-based detection link,and the measured satellite images and simulation data sets of Jilin 1 and other satellites were taken as the test set.Ablation experiments were carried out on the cross-scale connection pyramid module and the dynamic anchor learning method module in the proposed method to verify the effectiveness of the modules.In addition,the proposed method is compared with other typical algorithms(HOG+SVM algorithm,DPM algorithm and Faster R-CNN algorithm),and the average accuracy is taken as the evaluation index to verify the effectiveness and robustness of the proposed algorithm for different scene images.Experimental results show that the detection method presented in this paper can achieve better detection results in different airport scenes and cloud shielding scenes.