Research On Semantic Segmentation Method Of Water Navigation Scene Image Based On Polarization Imaging

The visual perception technology of unmanned surface vehicles is developed to improve the environmental perception ability of unmanned surface vehicles.It is the key and the cornerstone of the realization of unmanned surface vehicle technology.In the water navigation scene,extreme situations often occur,the image quality collected by visible light vision is pool,and the accurate visual perception of the unmanned surface vehicles cannot be achieved.In response to this,it is necessary to improve the semantic understanding ability of unmanned surface vehicles for harsh navigation scenarios.The climate environment of the water navigation scene is harsh and complex.The polarization,as an inherent property of matter,still has the advantages of good stability,strong anti-interference ability,and no accumulation of errors over time in the harsh navigation scene.A polarization camera array is designed to collect images of water navigation scenes with different polarization directions,the image sensor IMX274 is selected as the front end of the image acquisition device,and the Jetson TX2 is used as the data processing unit.Based on this,the polarization camera array is used to collect images of water navigation scenes with polarization information to achieve semantic segmentation,and a polarization semantic segmentation method based on polarization degree images using deep learning is proposed to improve the performance of unmanned surface vehicles in harsh navigation scenarios.The detailed research contents of this thesis are as follows:(1)Real-time acquisition and processing of polarization images with different polarization directions in the same scene.The polarization camera array with NVIDIA Jetson TX2 is designed as the data processing unit,and IMX274 is selected as the image acquisition front end to obtain the images of different polarization directions at the same time of the same water navigation scene.After image registration,the polarization degree images for the training of polarization semantic segmentation network are generated;(2)Analysis of polarization characteristics of water navigation scene.Based on the polarization principle,the polarization characteristics of the polarization images collected by the polarization camera array under different scene modules and different influencing factors are analyzed.Using the polarization characteristic distribution of water navigation scene to assist polarization semantic segmentation network to perform semantic segmentation of harsh navigation scene images;(3)Polarization semantic segmentation dataset arrangement and implementation of polarization semantic segmentation model.Aiming at the shortcomings that conventional image segmentation cannot be refined in harsh navigation scenes,a semantic segmentation network combining polarization technology and deep learning is proposed.The data of four channels includes the RGB image of the original scene without the front polarizer and DOP image are input into the network of this thesis as a dataset,the final segmentation results are obtained,and the segmentation results are analyzed and compared.The experimental results show that the polarization semantic segmentation network proposed in this thesis has better segmentation effect than the traditional methods when segmenting the images of harsh navigation scenes,and the detection accuracy is also improved.This method can effectively improve the water perception ability of unmanned surface vehicles,has a wide application prospect,and is of major importance for the research of intelligent ships.