| Due to the narrow coastline and wide territorial sea area of our country,remote sensing images play an important role in showing geographical form.If object detection are used for coastline fortification monitoring,valuable information can be mined,which may support counter-terrorism,disaster prevention,marine rescue and so on.However,remote sensing images are different from natural images,mainly reflected in its different perspectives and occlusion.Therefore,traditional detector have high false alarm rates and poor robustness in ship detection and it is slow and inaccurate,so in the paper we aim to improve the one-stage object detection performance in remote sensing images.The main work and innovations of this paper are summarized as follows:(1)This paper presents the improved network based on YOLO v3 structure in Chapter 3.The original network Darknet is replaced by the improved Inception and ResNet module.The test results show that it improves performance by 16.5%compared to the original network and 4.5%higher than Faster R-CNN network.It could not only overcome the poor recognition effect in small ship,but also achieved a better test result in dataset named VisDrone-2018.(2)On the basis of SSD structure,we propose two different Fused Feature Map SSD(FFSSD)structure in Chapter 4.One is ’element-sum’,the other is ’concatenate’.The results of both methods show that the FFSSD algorithm of adjacent feature fusion is 8.6%higher than Faster R-CNN algorithm and 10.9%high than the original SSD method.(3)Decay NMS is introduced to model.In the detection of small ship object,the number of boxes to be screened is often large and arranged densely,and their confidence scores sometimes inhibit each other.In order to avoid this problem,the detection confidence of the close ship small object is attenuated according to a certain rule.The performance of the improved NMS can be improved by 1%on the original Yolo V3 and 10%on FFSSD structure based on adjacent feature fusion. |
PDF Full Text Request