Research On Ship Detection In SAR Images Based On Attention Mechanism And Feature Enhancement

Synthetic Aperture Rader(SAR)technology has been widely used in military and civilian fields due to its unique characteristics of all-sky,all-weather imaging.As the main carrier of Marine information,ships can collect and transmit all kinds of Marine information.It is important to locate and identify ships all day and all weather in the Marine field.As a result,SAR image ship detection has received widespread attention from scholars and has become one of the research hotspots inside and outside the country.It can monitor ships at sea in an all-round manner,optimize ship scheduling to improve transportation efficiency,and optimize ship route planning to avoid potential collision risks.This means that SAR ship detection plays an extremely important role in ship monitoring management,ship scheduling management and ship traffic management.The research on SAR ship detection has made remarkable achievements after years of development.However,there are still some problems in the aspects of ship feature extraction,feature enhancement and feature expression.The first problem is that speckle noise in SAR images will cover up the details of ships.When feature extraction is directly performed on the original image,the extracted ship features will contain many redundant information,lack of texture information and other details.This makes the extracted ship features ineffective for ship detection.The second problem is that ships in SAR images are often interfered by complex background clutter in inshore scenes such as wharfs and buildings,resulting in non-significant and blurry ship features.This interference makes the extraction of ship features inadequate and poses great difficulties for ship detection.The third problem is that the features of ships in SAR images are relatively simple,and they bear some resemblance in appearance to other objects,such as floes and reefs.It is also a difficult problem to accurately identify ships from these objects.The extraction of high-quality ship features forms the foundation for accurate ship positioning and identification.This thesis focuses on optimizing the ship feature extraction,and enhancing the salience and recognizability of ship features.Firstly,it explores pixel-level feature enhancement,addressing two key aspects: how to suppress noise interference and enhance the saliency of ships.Subsequently,the study incorporates context feature enhancement to further enrich ship feature representation and enhance the recognizability of ship features,ultimately leading to improved accuracy in ship detection.To address the first issue,this thesis presents a SAR ship detection method with denoising and feature refinement.Firstly,a denoising module is designed to suppress speckle noise.Then,a hierarchical feature fusion module is introduced to prevent dilution of low-level features by high-level features with semantic information,retaining more spatial positional information,and effectively improving the detection accuracy of small object ships.Finally,the feature refinement module is proposed to highlight the effective features,further enhancing the ship positioning accuracy.This method can be applied in ship monitoring management,enabling high-precision,all-weather,and all-time ship positioning.It contributes to the enhancement of ship monitoring capabilities and ensures intelligent ship management.To address the second issue,this thesis proposes a SAR ship detection method combining salience regions extraction and multiple branch attention.Firstly,a pre-screening module is designed to separate regions that may contain ships from complex inshore backgrounds,filtering out irrelevant background information.Then,during feature extraction,a multi-branch attention module is employed to enhance the saliency of ship features,making them more recognizable.Since port berth space is limited,the timely and accurate determination of ship positions is crucial for providing clear guidance and planning for ship berthing,preventing congestion,and improving cargo loading and unloading efficiency.This method facilitates the orderly berthing and cargo handling of ships,enabling efficient ship scheduling.To address the third issue,this thesis presents a SAR ship detection method based on feature enhancement and contextual information.Firstly,it utilizes a feature extraction module with receptive field weight regularization to extract features with varying receptive field sizes,enhancing the model’s feature extraction capabilities.Next,it enriches ship features by fusing contextual information,thereby strengthening feature representation.Finally,it employs a decoupled head based on CIoU to separately perform classification and localization tasks,making efficient use of feature information.It guides the prediction of the regression branch using the CIoU loss function,ensuring a closer fit between predicted and ground truth bounding boxes,ultimately yielding improved detection results.This method combines ship context information,evaluating collision risks between ships and optimizing ship route planning.It plays a crucial role in ensuring safe ship navigation and has significant practical value in ship traffic management.The thesis conducts a comprehensive analysis and experimentation on the methods outlined above.Comparative experiments with current mainstream methods demonstrate that the proposed methods yield the best detection results.This validates the effectiveness of the proposed methods.This research is of utmost importance in advancing the practical application of SAR image ship detection technology in ship monitoring management,ship scheduling management,and ship traffic management.