Research On Key Technologies For Ship Target Detection In High Resolution Optical Remote Sensing Images

The sea ships,like aircraft carriers and warships,are charged with performing major military tasks such as maritime reconnaissance,defense and combat.Therefore,the ships are important carriers of military tactics and strategic objectives.The ability to detect ship targets accurately and efficiently is the foundation for unmanned aircraft to implement the target attack and the battlefield situation assessment,and it can also significantly improve the precision strike capability of existing weapons.Thus,it is necessary to studying the ship target detection.This paper focuses on the detection of ship targets in optical remote sensing images with high resolution.In order to accurately to detect the ship targets under complex harbor environment,we mainly study the sea-land separation,the extraction of line segment,the oriented region proposals and the convolutional neural network.This paper mainly focuses on the following four aspects:1.The sea-land separation based on structure saliency and weighted mean shift is proposed.First,a land membership function which utilizes edges and corners to measure the geometric structure saliency is proposed,and it is able to accurately calculate the land membership of most pixels.Then,in order to meet the requirements that the homogeneous region with similar color should belong to the sea or land areas,the input images are divided into homogeneous regions by using the weighted mean shift.Next,the Markov Random Field is employed to combine the land membership and the homogeneous regional consistency principle into an energy function.The Graph-Cut algorithm is then used to efficiently solve this function to obtain the separation of sea and land.A large number of experiments show that the sea-land separation method can correctly separate the ocean pixels from the land area and keep their boundaries accurate,which can provide the regions of interest and context information for the subsequent ship target detection.2.A line feature extraction method for remote sensing images based on the line probability map is proposed.Firstly,a line probability map based on the structure tensor is presented.The probability map not only takes advantage of the gradient amplitude information commonly used in the traditional methods,but also exploits the consistency of the gradient phase around each point,which is helpful for the line detection in the weak contrast images.Then,a series of edge chains are obtained by using the edge drawing algorithm from the line probability map.In order to make the detected line segments more complete,these edge chains are further split and merged.Next,they are quickly cut into candidate lines by using the least squares.Finally,we apply the gradient phase and amplitude criterion based on Helmholtz to validate these detected line segments.This can detect more accurate and complete line segments,and reduce the generation of the false line segments,which is beneficial to analyze the geometrical structures in remote sensing images,and detect ships and other targets.3.A ship detection method via ship head classification and body boundary determination is proposed.Firstly,we generate novel ship head features in the transformed domain of polar coordinate,where the ship heads have an approximate trapezoid shape and can be more easily detected.Then,these features are used in the classification based on SVM to detect the ship head candidates,and give the important information of initial ship head direction.Next,the surrounding consistent line segments are utilized to refine the ship direction,and the ship boundary is determined based on the saliency of directional gradient information symmetrical about the ship body.Finally,the context information of sea areas is introduced to remove false alarms.Experimental results show that the proposed method can accurately and robustly detect the warships in high-resolution optical remote sensing images.4.A ship detection and recognition method based on oriented region proposals and convolutional neural network is proposed.We first propose a ship oriented region proposals method based on region merging.According to the designed merger rules,it can generate more accurate and complete ship candidate areas,and provide the oriented minimum enclosing rectangles of the candidate regions,which can avoid the rotation problem in the subsequent recognition algorithm.Then,the candidate regions are sorted by using the relevant features of the ship.Our ship oriented region proposals method can ensure a higher recall rate when generating fewer candidate areas,which can reduce the computational complexity of the subsequent recognition algorithm.Next,we design the convolutional neural network to recognize the ship target.The feature extraction and recognition of ship target are effectively combined in the network,which overcomes the shortcoming that the traditional manual design features are hard to get a good balance.This network can also produce better ship recognition performance with fewer network layers and parameters.Our method can accurately divided the ship targets into aircraft carriers,warships or other types of ships,which can provide more effective information such as the location and type of ship targets for the target attack.