| The mission of ship detection in optical remote sensing images is an important task for national defense security.China has a vast sea area.Monitoring the sea channel and inland waterway by remote sensing can effectively improve the safety of water transportation and effectively improve the dispatching capability of port.With the continuous development of remote sensing technology,the resolution of remote sensing images is also constantly improving.The improvement of remote sensing image quality brings us more abundant image information,and it also increases the difficulty for our analysis of images.In this paper,the ship detection task is disassembled into two stages: coarse segmentation and fine segmentation.In the rough segmentation stage,land and water division is performed to shield the land portion.In the fine segmentation stage,on the basis of water and land segmentation,the ship targets is detected.Firstly,this paper implements water-land segmentation in the coarse segmentation stage.Based on the classical semantic segmentation network SegNet,this paper proposes an improved SegNet_plus network.The improvement direction is as follows: Improvement 1: A high-resolution input layer is designed in SegNet_plus which makes the input image have more abundant details.Improvement 2: This paper proposes the Encoder Block and the Decoder Block structure.The common convolution and dilated convolution two-way extraction feature is used in the Block,The characteristics of the object itself and the features of the surrounding environment are comprehensively utilized,so that the features extracted by the network are more abundant.Improvement 3: Based on the original residual connection of the SegNet network,the residual connection is added in the Encoder Block and Decoder Block structures to reduce the gradient disappearance during network propagation.The algorithm proposed in this paper achieves a correct rate of 97.33% in the self-built data set,which is higher than other algorithms compared in this paper.Then,in the fine segmentation stage,based on the target detection network SSD,the improved SSD_plus network structure is proposed.The improvement is as follows: Improvement 1: The high-resolution input layer is designed to reduce the information loss of the object,and the structure of the network is adjusted according to the size of the input layer,which deepens the network and improves the ability of the network to recognize objects.Improvement 2: Using the K-means algorithm to cluster the ship targets of the self-built data set which makes the aspect ratio setting of the default box more reasonable.Based on the above improvements,the detection rate of the SSD_plus network on the self-built data set is 71.44%,which is higher than other algorithms compared in this paper.In addition,this paper proposes a two-stage detection method,namely SegNet_plus+SSD_plus.When detecting a ship,firstly generate a land mask based on the result of land and water segmentation,and fine-tune the pattern processing such as hole filling and expansion is used.After shielding the land with a land mask,and use SSD_plus for ship detection.The mAP value of SegNet_plus+SSD_plus on the self-built data set is 74.50%,which is higher than the SSD_plus,and also validates the effectiveness of the SegNet_plus water-land segmentation network.Finally,this paper uses Python language and PyQt,Tensorflow,Tensorboard,MySQL and other development tools to integrate the SegNet_plus algorithm proposed by the land and water segmentation and the SSD_plus algorithm proposed by ship detection to design a ship detection platform system. |
