Detection Of Compound Object And Core Elements Of Airport In High Resolution Remote Sensing Images Based On Deep Learning

With the rapid development of satellite technology,high resolution remote sensing images are increasing day by day.Therefore,it is urgent to implement the automatic interpretation of remote sensing images,and object detection is an important research content in automatic interpreting of remote sensing images.In the context of depth learning,especially convolution neural networks,the thesis focuses on study the detection of airport,harbor,and the core elements in the detected airport using combination of deep learning features and handcrafted features.The proposed methods are integrated into the object detection platform for high resolution remote sensing images(ELU).The main work and innovation in this thesis are as follows:(1)As a compound object,the principal structure of the airport –runway,has obvious parallel line features,therefore,the method combing the extraction of airport interest area(ROI)using geometry and visual saliency map,and further classification of ROI by residual network is proposed to detect the airport.A local entropy saliency map is improved to combine with the geometric saliency map to determine candidate region of airport(ROI)in image.Then the ResNet(residual network)is used to discriminate the categories of the candidate area and obtain the airport detection results.Compared with traditional airport detection method,experimental results show that the proposed method improves recall and precision rates by 5.56% and 22.15%,respectively.Compared with end-to-end deep learning detection method,SSD(single shot multi-Box detector),our method improves recall and precision rates by 1.85% and 13.91%,respectively.(2)The harbor,as another compound object in our study,must appear at the sealand interchange,and there is no obvious primary and secondary relations between its composition structure –docks.Therefore,for harbor detection,a dual-resolution object detection method is proposed.ResNet is first used to determine the sea-land interchange in low resolution image,and then ResNet is used to detect docks in higher resolution image.The adjacent docks are merged to obtain final harbor detection results.Compared with the single-resolution ResNet method,experimental results show that in the case that the recall rate remains about the same(up to 95%),the precision rate of our harbor detection method increased by 11.06%,and time efficiency increased by 27.91%.(3)For airplane detection,based on the fact that airplane objects generally are very small,an improved SSD network based on multi-scale deep learning feature is proposed to improve the detection result for small objects.Experiments show that the recall and precision rates of the improved network proposed in this thesis increase 5.88% and 3.66%,respectively,for small objects of size less than 32 × 32.(4)For airport runway detection,according to fact that runway is the major structure of airport,a method based on airport contour for runway detection is proposed.ResNet is first used to obtain accurate airport area,then saliency map and binary algorithm with 0 value correction is used,by further removing small connected area and filling hole operations,finally the airport contour is obtained.On this basis,the runway can be quickly detected based on parallel lines within airport contours.Experimental results show that the proposed method in this thesis can maintain strong robustness while gray value of airport is close to surrounding environment.(5)The airport shelters are generally located at the end of runway and taxiway connecting area.For airport shelter detection,a method based on adversarial sample for airport shelter detection is proposed because of fewer airport shelter samples.Through skeleton refinement and search of endpoint within airport contour,endpoint is extracted as the suspected shelter area.The adversarial samples are used to increase the number of shelter samples and train the network to determine final airport shelter detection results.Experimental results show that after adding adversarial samples,detection recall rate and precision rate increased by 3.90% and 12.35%,respectively.