Research On Infrared Imaging Ship Target Detection Under Marine Environment

As the dominant carrier of marine transportation,ships are the key targets of maritime monitoring.Ship detection in complex marine environments is of great strategic significance in military and civilian fields such as national defense construction,fishery management,navigation supervision,etc.Due to strong environmental adaptability and all-weather reconnaissance capabilities,infrared imaging detection system is one of the important detection methods in marine environment monitoring such as ship target detection.However,different from visible and Synthetic Aperture Radar(SAR)images,the development of infrared ship detection is relatively slow because of the lack of public infrared ship datasets.The main challenges for infrared ship detection include low image resolution,single-channel information,complex ocean environment,small scale and weak semantic information of ship targets.Aiming at the problems above,this thesis studies the model-driven method,data-driven method and model-data joint-driven method,and completes theoretical research,dataset construction,and design of algorithms with high precision and robustness.The main research contents of this thesis are as follows.1.The fundamental theories of infrared ship detection are studied,including the basic theories of visual feature modeling,machine learning classifier,and deep learning theories such as backbone network and object detection framework.In-depth analysis of the infrared imaging characteristics and the visual characteristics of the target and background in the infrared image.Infrared imaging characteristics and visual features of targets and background in infrared images are also analyzed in depth.2.To cope with the lack of public datasets,an infrared ship detection dataset is established.This dataset contains 1284 infrared images and includes abundant offshore,near-shore,and inshore scenes,as well as complex weather such as heavy clouds and big waves.3.Aiming at the problems of land interference,complex environment and various false alarms,an infrared ship detection method based on visual feature modeling is proposed.This algorithm utilizes a sea-land segmentation algorithm based on superpixel segmentation to remove land areas,constructs a visual feature space according to the characteristics of infrared ships,and combines random forests to conduct ship detection.Experiments demonstrate that the false alarms in land areas are reduced efficiently,and the detection performance is satisfactory in most scenarios.4.Considering the slow speed and poor robustness of model-driven methods,a two-stage infrared ship detection model based on improved Faster RCNN is proposed.An adaptive feature fusion network is introduced to balance the semantic information and position information of different feature layers,which reduces missed detection of small targets.A pixel-level spatial attention network is proposed to enhance target information while suppressing clutter in complex environments,effectively reducing false alarms under complex scenes.The feature extraction module in prediction head is adjusted to enhance the feature expression ability,achieving high detection performance under few-shot learning.5.A model-data joint-driven context perception network is proposed,which combines the advantages and complements the disadvantages of the model-driven and data-driven methods.In this network,a receptive field expansion module based on dilated convolution constructs balanced local and non-local features,effectively reducing the missed detection of extremely small targets;a contextual attention network is proposed to enhance the information of the target itself and the context area,providing complementary information for small targets with weak semantics;a knowledge-driven prediction head is designed to integrate visual features into the neural network through a supervised learning branch,so that prior knowledge can be backpropagated to the entire network.Extensive experiments demonstrate that the proposed network achieves state-of-the-art performance on infrared ship detection dataset.