Combined Supervised And Semi-Supervised Learning For SAR Ship Detection

Synthetic Aperture Radar(SAR)is an active microwave imaging sensor capable of generating high-resolution images in all-weather,all-day conditions and is widely used in SAR ship detection.Traditional SAR ship detection methods rely on manually designed features,which makes it difficult to guarantee detection performance in complex scenarios.In contrast,deep learning methods can automatically learn the feature representation of the target and gradually become a hot research direction in the field of SAR ship detection.Not only that,the emergence of semi-supervised learning methods has further promoted the development of deep learning algorithms,such as student-teacher networks,which enable deep learning models to achieve better performance despite insufficient labeled samples.The ship detection task in SAR images is challenging due to the large variations in ship scale and complex backgrounds,which necessitates the model with good robustness and feature extraction capabilities.Supervised learningbased anchor-based methods commonly use feature pyramids to extract multiscale features and output bounding boxes to achieve object detection.However,these methods neglect the fusion of adjacent layers,thus limiting the model to obtain richer semantic spatial information.In addition,in the case of densely arranged or rotated targets,it is easy to cause problems such as box overlap and inaccurate localization.The aforementioned data-driven supervised approach is capable of achieving high detection accuracy,but depends on a large number of pixel-level labels for support.To alleviate this problem,semi-supervised learning methods are widely used.Traditional semi-supervised learning methods utilize consistency regularization and pseudo-labeling to leverage unlabeled samples.However,during the training process,these models may encounter issues such as coupling and error accumulation,which can ultimately limit detection accuracy.To address the aforementioned issues,this paper proposes improvements in model structure,feature optimization,and algorithm flow.The main contributions are summarized as follows:1.A ship detection method based on adjacent context-guide fusion module and dense weighted skip connection is designed: the spatial density map method is used to transform the input labels into the spatial density map,and regress the ship in the pixel dimension,and utilizing a clustering algorithm to obtain the rotatable bounding box;adjacent context-guide fusion module is designed to extract strong semantic information of adjacent high-level features and fuse them with low-level features in the channel dimension,thereby improving the fusion of adjacent layers;the dual-path enhanced pyramid is designed to learn the pyramid feature representation through two paths of enhanced information flow,allowing for the extraction of rich semantic and spatial information,which improves the detection of multi-scale ships;dense weighted skip connection is designed to fuse features of all scales in cross-layer connections and weighted fusion using trainable weights.This enriches the feature space of the decoder and focuses on critical features.2.A ship detection method based on adversarial cross student-teacher network is designed: two student-teacher networks were designed so that the two sets of student networks could learn independently while cross-constraining with pseudo-labels generated by the teacher network.This approach helps to avoid coupling and error accumulation between the networks;the feature discriminator is designed to form an adversarial relationship between the student-teacher network and the discriminator,so that the features extracted by the student network are consistent with the teacher network and result in feature constraints;the edge disturbance module is designed to adding additional edge interference,which allow the network to extract more robust edge information;the feature refinement module is designed to modulate the extracted features and enable the network to focus on crucial channels and features.Through iterative training,higher quality pseudo label and robust student networks are gradually obtained.Experiments are conducted on SSDD,AIR-SARShip-1.0 and HRSID public ship detection datasets,which demonstrate that the proposed method outperforms the mainstream methods in terms of accuracy and robustness.In conclusion,we summarize our work and provide an outlook on further research progress.