Research On Satellite Ship Image Recognition And Semantic Segmentation Methods

With the development and utilization of the ocean by humans,the ocean has become increasingly important to humans.Therefore,the implementation of marine vessel testing plays a vital role in national economic development,maritime traffic safety,rational use of marine resources and territorial security.As an emerging technology,the ship identification technology based on optical satellite image provides a new technical means for long-distance maritime dynamic monitoring system.Compared with the traditional detection system,it has many obvious advantages such as remote and wide-area dynamic monitoring,short monitoring period and high recognition rate.Therefore,it is very important to perform satellite ship image recognition and segmentation quickly,efficiently and accurately.This paper focuses on the neural network structure RIRnet for satellite ship image recognition.Firstly,from the design of the network model,comprehensive research is carried out from the aspects of model complexity,the disappearance of gradient propagation and the characteristics of the data set itself.The RIRnet network structure is structurally divided into three parts:the initial layer,the transition layer,and the output layer.The RIRnet network structure designed in this paper has three characteristics:(1)It mainly adopts the structure of residual residual embedded residual as the basic structure to make gradient propagation easier;(2)The initial layer uses five different convolution kernels from different Dimensions extract features and feature blending.(3)The transition layer also uses the residual embedding residual structure constructed by two different convolution kernels,and splicing the residual blocks of the two convolution kernel constructs to make the features better.Finally,experiments were carried out on satellite ship image data and compared with other classical neural network models.Secondly,for image segmentation,the U-Net network and the improved U-Net are used to semantically segment the satellite ship image.Both networks use the deconvolution structure to achieve upsampling.The U-Net network structure is coarse for the target edge segmentation,and the segmentation accuracy of the small target image is poor.Therefore,the loss function and structure of the U-Net network are improved,and the RIRnet network is embedded into the U-Net structure to achieve better feature extraction.Finally,the verification is carried out on the satellite ship image dataset.The improved U-Net network segmentation effect is better,and the target edge detail and small target segmentation ability are stronger.It proves that the improvement of U-Net has achieved certain effects.