Research On The Method Of Identifying Underwater Monuments Under The Constraint Of Perception

In the process of performing underwater monument recognition,the monuments are immersed underwater for a long time causing problems such as rust stains and adhesion of marine organisms,resulting in the original form of the monuments that cannot be directly extracted by the AUV,thus reducing the accuracy of the recognition algorithm.Further,it is difficult to segment the underwater monuments from the surrounding environment due to the cluttered background of the collected underwater images and insufficient underwater light,which causes the algorithm to miss detection.Most of the underwater monuments are only wrecked after the corrosion of seawater causing the target image features captured by AUV to be missing,which affects the accuracy of target detection.Meanwhile,the underwater archaeological data is not easy to obtain leading to little training data,resulting in poor generalization performance of the recognition algorithm.In response to the above practical problems,this paper has done a lot of research.Specifically,the main contributions of this paper are as follows.(1)An underwater marine life attachment monument identification network is proposed.Firstly,Res Net50 is used as a feature extractor to extract features from the input underwater monument images,and the high-level features of the extracted underwater monuments are introduced with channel attention and spatial attention to initially localize the underwater monuments,and then the local areas containing the monuments are finely searched to achieve accurate localization of the monuments covered by marine organisms.Then,the invisible parts of the monuments attached by marine organisms are restored in two stages,the features are restored by over three hourglass models,the a priori information is used to predict the invisible parts of the monuments,and the attention module is added to capture the contextual information according to the different postures of the monuments,so as to restore the parts of the underwater monuments.Finally,the features from the spatial distribution search module and the underwater marine organism attachment feature recovery module are fused,and then the bounding box regression is performed.To verify the effectiveness of the algorithm,simulations are performed on a homemade dataset and three sets of simulation experiments are conducted under the conditions of normal water environment,monuments attached by marine organisms and high similarity to the background.The experimental results show that the proposed algorithm in this paper improves the m AP by 1.67%,2.93% and 1.63% in the three sets of experiments when compared with the advanced algorithm,respectively.(2)Underwater Incomplete Target Recognition Network via Generating Image Features is proposed.Although MLAMNet has improved the accuracy of underwater monuments recognition to some extent,the missing features of underwater monuments and the small training data have caused some impact on the accuracy of the algorithm.To address the above problems,firstly,feature generation is carried out for targets with incomplete features by means of a dual discriminator and a generator,which are divided into two sub-modules,the former one is responsible for generating features and the latter one for noise reduction,which reduces the noise generated by the generator while generating features and improves the accuracy of the algorithm.Then,FPN is used to fuse multi-level features and thus extract regions of interest.Finally,supervised contrast learning is introduced to few-shot learning theory,and contrast proposal coding is used for few-shot target detection,and contrast branches are added to the region of interest features to improve the intra-class similarity and inter-class distance of the target and enhance the generalization performance of the algorithm.To verify the effectiveness of the algorithm in this paper,a UIFI dataset is produced.The experimental results show that the m AP of the algorithm in this paper is improved by 0.86% and 1.29% under the interference of poorly lit and semi-buried monuments,respectively,and the m AP for shipwreck identification reaches the highest level under all four sets of experiments.