Ship Detection In Multi-scale Remote Sensing Image Based On Hierarchical Depth Network

Synthetic aperture radar,SAR for short.This system can observe the earth 24 hours a day regardless of weather and climate.It can also be called an "all-weather,allweather" radar observation system.In actual use,it is usually carried on aircraft or satellite to achieve earth observation.The system is widely used in the military and civil fields.In the military,it can monitor the marine environment and interests of the marine land.In the civil,it can be used for port traffic management,and it can also be used for illegal behavior monitoring such as smuggling.In recent years,various countries have invested a lot in SAR image ship detection.However,the large field of view SAR remote sensing image contains complex background information,and there is more false alarm interference.In addition,the ship target also has significant scale differences.This paper uses deep learning methods to detect ships in SAR remote sensing images.Aiming at the related difficulties in SAR remote sensing ship detection,this paper proposes a multi-scale remote sensing image ship detection technology based on a hierarchical deep network.The main research work and results are as follows:This article is aimed at the characteristics of SAR remote sensing images.Ships existing in the ocean area,and land scene information redundancy during detection,which is prone to false alarm interference.First,extract the location information of the ocean and land,and use the location information to guide the false alarm elimination in the detection stage.A method for extracting sea and land scene information based on depth separable convolution is proposed.By replacing the standard convolution in Deeplab V3+ with depth separable convolution.Carry out sea and land scene information extraction.Aiming at the large difference of ship scales in SAR remote sensing images,this paper proposes a multi-scale ship detection method based on recursive feature pyramid.Based on Efficient Net.Through compound scaling,the value of the three parameters that balance the network depth,width and resolution is obtained,to obtain a better detection model.In view of the multi-scale,the feature fusion module introduces a recursive feature pyramid.It can enhance network feature characterization capabilities and improve the accuracy of ship detection.This article focuses on the characteristics of ships docking: ships in different directions along the port and ships may be densely arranged.For the description of the horizontal,it is easy to include redundant information about the port and land.Using the horizontal frame to return in the dense area of ships,can easily cause the return frame to overlap and lead to missed detections.Based on the problems,this paper proposes a refined rotating frame regression method.This method first adds angle information to obtain the rotating frame.Then use the feature refining module,through bilinear interpolation,the problem of inaccurate descriptions of overlapping boxes is solved.