Research On Underwater Occlusion Image Reconstruction And Obejct Recognition For AUV

At present,the gradual globalization of the world economy and the increasing energy demand.The exploration,development and utilization of deep-sea resources are increasingly emphasized by many developed countries in science and technology.As a significant field of high technology,the development of ocean science and technology is promoted intelligent underwater machinery and equipment to replace human resources to achieve autonomous marine protection and management.AUV is created as an intelligent detection device that can operate autonomously underwater.AUV uses onboard sensing systems to sense the surrounding environment.As an important part of the perception system,the underwater visual perception system has been widely valued by scholars for its advantage of directly acquiring environmental visualization features similar to the human visual system.However,underwater image blur,interference,foreign body occlusion and other phenomena result in incomplete target features.And the scarcity of underwater data samples is difficult to meet the demand of data for deep learning.This has raised a great challenge for AUV to accurately recognize underwater target.To solve the above problems,this paper proposes several underwater occlusion image reconstruction and target recognition models based on the deep learning framework to achieve accurate recognition of underwater occlusion targets.The main studies are as follows:(1)The fine restoration method of the incomplete image with external features and image features is proposed.Firstly,the DMN+ algorithm is used to combine the incomplete image and the related external features Then generate an optimized image of the defective image containing external features and image features.Secondly,the generative adversarial network with gradient penalty constraints is constructed to guide the generator to perform coarse restoration of the optimized mutilated images to obtain the coarse restoration map of the target to be repaired.Finally,the coarse restoration map is further optimized by the relevant feature coherence module to obtain the fine restoration image.(2)The underwater fine reconstruction method based on environmental feature fusion is proposed.First,the image significant feature extraction module is constructed based on contrast learning to obtain the significant features of the image.Secondly,constructing a hierarchical retrieval module for environmental feature noticing mechanisms to obtain information on environmental features with relevance.Finally,the coarse-to-fine image reconstruction model with gradient penalty constraints is constructed based on the WGAN network to generate finely reconstructed images.(3)The underwater occlusion target recognition method based on the fusion of significant environmental features is proposed.First,the target significant feature extraction module and the environment feature attention module are constructed respectively to obtain the significant features of the target and the related significant environment features.Secondly,the construction of a comparative graph structure based on GNN for the significant features of the target and the relevant environmental features.Finally,the interaction of target and background in the contrast graph structure is used to construct the underwater occlusion target recognition model to achieve accurate recognition of the occlusion target.(4)The two-stage image reconstruction strategy based on an underwater occlusion target recognition method is proposed.Firstly,the original underwater image is preprocessed by using the proposed underwater incomplete image restoration model and underwater fine reconstruction model.Second,constructing a target recognition model with feature adaptive boundary regression to compensate for the ambiguity in the real bounding box caused by the uncertainty of the labeling process.Finally,the constructed two-stage image processing strategy is used to achieve reconstruction and recognition of underwater mutilated targets.(5)Systematic validation experiments of the proposed algorithm modules are carried out based on the laboratory hardware and software platform.The experimental results show that the proposed method achieves optimal performance in all scenarios compared with existing representative methods.