Research On Mid-low Resolution Remote Sensing Images Ship Detection Method

China has a vast sea area and rich marine resources.The research on marine remote sensing technology is of great significance for China’s sea area monitoring,grasping the initiative of Maritime War,protecting people’s life and property safety and National Territorial Sea sovereignty.In the military field,the use of satellite remote sensing technology can quickly and efficiently monitor sea ships,respond to potential external threats in time,and protect China’s long coastline.In the civil field,the rapid detection and recognition of ship targets is also the key technology for the application of sea traffic control,port management,maritime rescue and combating illegal smuggling.Therefore,sea surface ship target detection based on satellite remote sensing images has very key research significance and high research value.However,due to the vast sea area,the distribution of ship targets on the sea is extremely sparse,and the computing resources are limited.At present,satellites in the field mostly obtain large width medium and low resolution images in the sea area to ”scan” the sea area.Therefore,this subject mainly studies the ship target monitoring method of medium and low resolution optical remote sensing images.The difficulties of ship detection in medium and low resolution optical remote sensing images can be summarized as follows:(1)the detection environment is complex,and the land area,thick cloud,light and shadow,wake and ocean waves in the image increase the difficulty of ship target detection;(2)The target size is small,the ship target size in medium and low resolution images is small,and the feature extraction is difficult,which increases the difficulty of ship target detection;(3)The wide range of remote sensing image and sparse distribution of ship targets restrict the efficiency of detection algorithm.In view of the above difficult problems,how to extract ship targets quickly and accurately is the focus of this research.Based on Sentinel-2 satellite,this paper constructs a large width low and medium resolution remote sensing image data set,and constructs a two-stage ship target detection framework for the difficulties of ship target detection in low and medium resolution remote sensing images.Firstly,the framework extracts a wide range of suspected target areas,and then detects ship targets in the extracted target areas,so as to improve the detection efficiency and accuracy.In addition,the effectiveness of the proposed framework is verified in the self built sentry-2 satellite remote sensing image data set.The main works of the project are as follows:(1)Aiming at the problems of large size and sparse distribution of ships in medium and low resolution remote sensing images,a suspected target region extraction method based on semantic segmentation and gray feature modeling is proposed.Firstly,the sea area is extracted by semantic segmentation Segformer algorithm to filter the land,thick cloud and other areas in the remote sensing image.On the one hand,the extraction of sea area improves the efficiency of detection algorithm,on the other hand,it reduces the existence of false alarm in land and other areas.Then,the gray characteristics of the extracted sea area are modeled,the gray characteristics of the target area and non target area are analyzed,and the non target area without ship target is filtered out.Finally,the reserved suspected target area is sent to the subsequent detection network.(2)Aiming at the problems of small ship target size and loss of detail information in medium and low resolution remote sensing images,this paper introduces image superresolution into the field of target detection,and proposes a new ship detection network based on image super-resolution(ShipSRDet algorithm for short).ShipSRDet algorithm uses image super-resolution to restore the detail information in the image,and sends the super-resolution image and the information rich features generated in the super-resolution process to the target detection module,so that the detection module can more effectively extract the appearance features of ship targets to achieve more superior detection performance.Experimental results on HRSC2016,DOTA and NWPU VHR-10 datasets verify the effectiveness and generality of ShipSRDet algorithm.(3)Aiming at the problem of ship target information loss in low and medium resolution remote sensing images,this paper introduces knowledge distillation and further proposes a low and medium resolution ship target detection method based on FDNet based on ShipSRDet algorithm.The FDNet algorithm framework includes two parts: teacher network and student network.The student network includes super division module and detection module.The super division image output by the super division module is input to the detection module for detection;The teacher network is a detector,and its network structure is completely consistent with the detection module of the student network.The input of teacher network is high-resolution remote sensing image,while the input of student network is medium and low resolution remote sensing image.FDNet algorithm transmits the rich information extracted by the teacher network to the student network,so that the student network can extract more features that have gain for detection from low and medium resolution remote sensing images,so as to improve the detection performance of the student network.The experimental results show that FDNet algorithm can improve the effective feature extraction ability of student network,and then improve the performance of ship target detection.