Optical Image Restoration And Semantic Segmentation Algorithm Of Unmanned Surface Vehicle In The Sea Environment Under The Hazy And Rainy Weather

In response to the policy of building a country with strong military power on the ocean,the research of Unmanned Surface Vehicle is of great significance.As the small ship,because of the characteristics of flexible maneuverability,large range of activities,long continuous operation time and low construction cost,Unmanned Surface Vehicle has gradually become the focus of research at home and abroad.The perception ability of Unmanned Surface Vehicle is the basis for the operation of Unmanned Surface Vehicle.The pixel-level semantic segmentation of the sea scenes can support the Unmanned Surface Vehicle at a deeper level to realize accurate identification of passable areas and the positioning and recognition of targets in the sea.However,when the Unmanned Surface Vehicle encounters complicated sea conditions such as rainy and hazy weather,its perception ability will be severely reduced so that the surface Unmanned Surface Vehicle can not perform its tasks normally.Therefore,it has important practical application value to carry out the research of image dehazing,deraining and segmentation algorithms for Unmanned Surface Vehicle in complex weather.Relying on the "QZ" Unmanned Surface Vehicle experimental platform,this paper carried out the research on the optical image restoration algorithm and the sea scene semantic segmentation algorithm for Unmanned Surface Vehicle in the rainy and hazy environment.Through the deep learning network,the images of the hazy and rainy weather are restored into the high-definition images.In order to improve the perception of the environment by the Unmanned Surface Vehicle,the semantic segmentation algorithm of the sea scene is on research.The main contributions are as follows:1.Through the research,it is found that the appearance of the haze in the image is complex and difficult to remove.This paper proposes a GAN-U-Net++ dehazing network.The network improvement generator U-Net++ network in the layer-to-layer connection method improves the utilization of the sea surface information of the image.The coordinate attention mechanism is proposed to improve the edge sharpening ability of small targets on the sea surface in the process of dehazing and the clarity of the targets on the sea.2.A variational auto-encoder is used to generate the simulated real scenes as the rain dataset.Combined with the density,direction and distribution similarity of rain lines on multi-scale images,a pyramid feature extraction and fusion rain removal network is proposed.Among them,the different numbers of coarse feature extraction and fusion modules can constitute light-weight deraining network or the accuracy of the deraining network.3.Through the research of texture feature information statistics and residual network,the semantic segmentation algorithm for texture information learning is proposed.The basic structure of the vector texture statistical operator and the adjacent texture statistical operator constitutes a texture information enhancement module and a multi-scale texture information extraction module.It can improve the semantic segmentation ability of the sea scene.