Research On Ship Detection Based On Few-shot Learning For SAR Image

Ship target detection is of great significance for safeguarding our country’s territorial security and legal rights and promoting the healthy development of fisheries.SAR image has become an important data source for ship target detection due to its excellent image characteristics.With the development of synthetic aperture radar,the amount of data continues to increase,and traditional algorithms cannot achieve real-time detection of ship targets.Despite the detection algorithms based on deep learning can achieve real-time detection,such algorithms are mainly for optical images,and the detection effect on SAR images is not ideal.Moreover,the target detection algorithm of deep learning requires a large number of labeled samples,and SAR images have fewer data and are difficult to label compared to optical data.In response to the above problems,this essay studies the algorithm of ship target detection in SAR images based on few-shot learning.The main research content of the article is as follows:(1)The characteristics of SAR image and optical image are quite different.The algorithms which can obtain better detection results on optical images maybe not achieve the same results on SAR images.Aiming at the problems of SAR image being easily interfered by background clutter,the image contains less information and the scale of ship target is small,this article improves the YOLOv3 object detection algorithm,which mainly includes the following four aspects: 1)Use K-means clustering algorithm to reselect the size of the anchor box;2)Improve the residual structure to enhance the ability of the network to extract features;3)Use the volume integral solution idea to lighten the network and improve the detection speed;4)Optimize the function used to calculate the loss.Experimental results show that the improved YOLOv3 is more suitable for the detection of ship targets in SAR images than before,and has higher detection accuracy under the premise of meeting the detection speed.(2)The improved YOLOv3 algorithm has achieved better detection performance,but this is inseparable from the training of a large amount of labeled data,and the generalization ability of the trained model is poor.When detecting new categories,a large amount of data still needs to be labeled.Therefore,this paper introduces a few-shot learning method,and on the basis of improving the feature extraction network Darknet-53 of YOLOv3 in Chapter 3,adding a feature weight adjustment module and a detection prediction module to construct a few-shot detection model.The model can learn meta-features from the base class,and only a few labeled samples for the new class can complete the detection.A large number of experiments have proved that the few-shot detection model constructed in this paper is better than other baselines,and has better detection performance and generalization ability.When the number of samples is small,the detection result is still similar to the result of the model trained with a large number of samples When the number of samples is small,the detection accuracy is still close to the model trained with a large number of samples.In addition,this essay also explores the training method of the model,and the influence of the number of base class samples on the detection results was analyzed.