Research On Sea Surface Ship Detection Based On High Resolution Remote Sensing Image

China is rich in marine resources.It is of great significance to develop the marine economy,maintain the national ocean,and build a maritime power.Ship is an important target for maritime traffic and resource detection.The detection of surface ships can be widely used in military and civil affairs.With the development of space remote sensing technology,the resolution of optical remote sensing satellites has also been continuously improved.It brings richer color and texture information for ship target detection.But it also brings some problems.For example,largely wide ship remote sensing image detection time is long.In a complex background,there are thin clouds,sea clutter,ship wakes,and the ship is parked in a port.These will lead to low detection reliability.Therefore,it is important to eliminate the interference from complex background.At the same time,fast and accurate detection of ship targets on the sea has become the focus of research in recent years.In view of the complex sea conditions,high-resolution ship remote sensing image has the characteristics of large width and rich details.This paper proposes a ship detection method,which mainly includes two parts: ship target candidate region location and ship target detection.Among them,the target candidate area uses saliency vision method.It solves the problem of long detection time of large and wide image.The saliency detection method can eliminate the interference of the cloud and fog.And it can quickly and accurately locate the ship candidate area.Ship target detection uses semantic segmentation method based on deep learning.It solves the problem of low detection accuracy caused by background clutter.The specific content includes:(1)High resolution ship remote sensing image has a wide range.It is easily affected by thin clouds and sea clutter.At the same time,it takes a long time to detect large and wide images.In this paper,a candidate region model of saliency detection based on adaptive superpixel segmentation is proposed.Firstly,the attraction function and the attribution function are used to adaptively form the superpixel.Then calculate the similarity between the image blocks and the connectivity value of the region.Significant results can be obtained by setting threshold.Finally,this paper proposes a segmentation method based on Sobel operator to binarize the salient region.The results are mapped to the original image and segmented to obtain candidate regions.(2)In view of the problem that the port background in the candidate region will still affect the detection results.This paper proposes a ship detection model based on deep semantic segmentation to further detect ships.It is based on Res Net.Firstly,the ship remote sensing image is roughly segmented by deep convolution neural network.Then,through the fully connected conditional random field,the model uses Gaussian pairwise potential and mean field approximation theorem.It establishes fully connected conditional random field as recurrent neural network for fine segmentation.In this way,an end-to-end connection is realized.It obtains the ship classification result.Through the candidate region location and ship detection method,the paper compares with other methods on the established dataset.The experiment shows that for different cases of high-resolution ship remote sensing image.The proposed method can classify ships accurately with clear edge contour,which is better than other models.The m IOU value is 84.3%.The time is 0.8s,which meets the requirement of high resolution remote sensing image of ship detection.