Research On Underwater Visual Perception And System Application Via Network Architecture Search

Underwater vision provides important perceptual information for close-range and highprecision marine operations.However,various degradation factors make underwater vision suffer from color distortion and low contrast,which in turn affect tasks such as visual exploration and accurate operations’ performance.Artificial intelligence perception technologies such as underwater image enhancement and object detection help underwater perception information mining.However,traditional hand-designed neural networks cannot meet the requirements of robustness and efficiency for underwater operations.To solve these problems and meet requirements of underwater tasks,this paper aims to create an efficient and high-performance underwater visual perception system through network architecture search technology.This paper designs a novel network architecture search scheme that can search network architectures for underwater scenes.It helps to improve the performance of underwater image enhancement and underwater object detection with clearer underwater images and more accurate detection results,based on which an underwater visual perception system is constructed.In addition,this paper uses multi-threading methods to optimize the processing rate of the system according to the hardware system,to meet the needs of the actual operation.This paper validates the effects of underwater image enhancement and detection through couples of experiments.And the system has been deployed in different robots and participated in underwater robot competitions for times,as a result it helps to win first prize in both humanmachine collaborative grasping and autonomous grasping tasks.In general,this paper designs the optimal network and deployment scheme for underwater tasks from the aspects of searching,inferencing,and deployment,which provides support on advanced underwater operation tasks for exploring and developing ocean with better assistance.