Research On Methods Of Maritime Object Detection Based On Self-supervised Representation Learning And Enhancement

Marine scene perception is a prerequisite for the effective completion of various tasks by unmanned vehicles and other maritime equipment,while object detection is a basic task in the field of scene perception and understanding.Therefore,it is of great significance to improve the accuracy of maritime object detection algorithm in the process of pursuing the improvement in the working ability of unmanned surface vehicles.In recent years,the deep learning-based object detection methods have made breakthroughs and achieved promising detection and recognition performance on large-scale standard datasets.However,due to the lack of maritime samples,the generalization ability of the models trained on the commonly used object detection datasets in the marine scene is quite weak.At the same time,because of marine environment is much more complex and changeable than its overland counterpart,it is difficult to obtain high-quality samples.Currently,the datasets available for maritime object detection are relatively small,along with imbalanced sample distribution.Therefore,the research on deep learning-based maritime object detection is still in the initial stage.In the context of the aforementioned factors,this thesis is committed to making full use of the existing marine datasets to improve the performance of maritime object detection,and further explores the application of deep learning technology in this field from the perspective of sea target feature learning and enhancement.The main points of this thesis are as follows:(1)Aiming at the problem that the existing maritime object detection datasets are small,a novel detection method based on self-supervised representation learning is proposed,which includes two stages:self-supervised ship feature learning and supervised maritime object detection.During the self-supervised stage,a comparative learning method is adopted to train the ship feature extraction model on a large unlabeled ship dataset,so that the ship features with high sample discrimination can be learned without category labels.Then in the supervised stage,the parameters of ship feature extraction model learned in the previous stage are used to initialize the backbone network of the detection model,and then the whole model is trained in a supervised manner on the maritime object detection datasets with both category and bounding box labels.The proposed method can make the maritime detection model have prior knowledge of the ocean scene to some extent and effectively improve the performance of the maritime object detection model.(2)To deal with the problem of class imbalance in the existing maritime object detection datasets,a ship detection method based on inter-class feature enhancement is proposed.Its core parts are a dynamic representation block of ship targets and a inter-class feature enhancement module.The ship target dynamic representation block is used to store the characteristics of various ship targets,which is initialized relying on the pre-trained maritime object detection model and dynamically updated in the subsequent training process.The feature enhancement module is used to adaptively learn effective auxiliary features from the dynamic representation block to enhance the features of ship targets.This method can make full use of the similarity among various types of ship targets,thus can improve the characterization ability of the model for a few types of ship targets,and further improve the performance of maritime ship detection significantly.