Research On Visual Recognition Technology Of Sea Surface Targets Based On Deep Learning

In recent years,maritime autonomous surface ships(MASS)and other marine unmanned equipment have become the development trend of the new generation of ships.In the face of complex marine environment,the ability of autonomous ships to detect and identify targets or obstacles is the primary condition for obstacle avoidance,planning and autonomous completion of tasks.Therefore,the environmental awareness and recognition system of autonomous ships needs to be real-time and accurate.Image segmentation and recognition technology based on visible light sensor and infrared sensor can obtain target feature information through rich image information and algorithms,but visible light sensor and infrared sensor have their own advantages and disadvantages and use limitations as separate recognition methods: visible light sensor has high resolution and color information,but it performs generally at night,rain and fog,etc;The infrared sensor has low resolution and no color information,but can be used around the clock.In order to better realize the sea scene perception,the fusion image recognition technology based on visible light and infrared sensors is the main development direction of sea target recognition technology.Aiming at the current situation of low accuracy of target perception and information extraction,complex sea environment,and difficult target recognition in autonomous ship navigation,this paper aims to achieve high-precision sky-obstacle-sea segmentation and sea target recognition,and focuses on the problems of complex weather conditions and many interference obstacles in sea environment perception,and carries out the research of visual recognition technology of typical sea targets based on visual sensors such as visible light and infrared,The segmentation and recognition of sea surface targets with high accuracy is realized.The main research contents and innovations of this paper are as follows:1.In order to solve the poor performance of existing semantic segmentation algorithms in the face of adverse weather conditions,small obstacles,sea reflection and other fuzzy visual conditions,this paper proposes a multistage feature aggregation and semantic feature separation network(MFASNet)for the real-time obstacle map estimation of unmanned ships,and proposes an efficient multistage feature aggregation(MFA)module to capture and fuse the different scale features and context information of the trunk at different stages,At the same time,a new loss function is designed to increase the separation between different semantic features and achieve more robust feature representation in highly diverse marine environments.These works effectively improve the real-time obstacle map estimation performance of unmanned ships.2.In order to solve the problems of existing target detection algorithms in the face of real ocean scenes,an algorithm based on YOLOv5 is proposed,which optimizes the input,loss function and detection frame of the depth learning network model.At the same time,the visible light sensor and infrared sensor are fused with depth information,and the pixel-level fusion,feature-level fusion and decision-level fusion are compared,forming a feature-level fusion algorithm based on VGG-ML network,It can accurately complete the ship detection task on the sea with high accuracy.