Research On Detection Algorithms Based On Deep-Learning Networks For Ship Objects In SAR Images

Synthetic aperture radar(SAR)is an active microwave remote sensing imaging sensor,which can obtain massive high-resolution and wide-scale remote sensing images.Compared with optical sensors,SAR has the all-day,all-weather,multi-angle and long-distance monitoring capabilities.With the continuous improvement of SAR imaging technology,it has been widely applied in all aspects of social and economic life,among which the automatic ship detection in SAR images has attracted the attention of scholars due to its important practical value in both civilian and military fields,and has become the current research hotspot.Therefore,it is very significant to study the task of ship detection in SAR images.Precise detection of specific objects in an image is one of the key research issues in the field of computer vision,whose core idea is to combine image processing technology and machine learning algorithm to accurately express and locate the objects in the input image.Traditional algorithms for ship detection in SAR images are mostly designed for specific scenes,which are highly dependent on predefined distributions or artificial designed features,and usually have poor generalization abilities.Recently,deep-learning algorithms have found wide applications in ship detection from SAR images for the reasons of data-driven,strong ability of feature expression,and high-accuracy of detection results,which thus have become the mainstream solutions in current object detection tasks.As a coherent imaging system,SAR inevitably generates speckle noises from the complex backgrounds in real scenes;The metal materials and the superstructure of the ships usually produce strong backscatters which will reshape the ship appearances in the SAR images and interfere with the detection process;Besides,ships are often clustered with large scale spans,and both shapes and forms of multi-scale targets in SAR images vary to some extent,most of all,small ships restrict deep networks to extract representative target features which makes the detection accuracy difficult to guarantee.In order to solve the above problems,this paper studies the ship detection in SAR images,focusing on the key technologies such as anchor-based and anchor-free detection networks,attention mechanism,and salient feature extraction,carrying out a series of works including problem analysis,theoretical research,network design,experimental comparison and verification,besides,a variety of different feature extraction and feature fusion methods are studied in this paper,and the proposed solutions are fully demonstrated through several public SAR datasets.The main research contents of this paper are as follows:An improved anchor-based model based on attention mechanism for ship detection in SAR images is proposed to detect multi-target ships with multi-scale sizes in front of complex backgrounds.Firstly,to achieve the best trade-off between detection accuracy and speed,we adopt the off-the-shelf YOLOv4 as the inspiration of our basic detection framework.Secondly,we design a thresholding attention module(TAM)that is embedded in very first layer of the network to perform denoising in the image-level.The TAM block can adaptively learn a set of thresholds according to the global information of the image to suppress noises,avoiding the invalid data flow of the network.Besides,in order to improve the detection performance of multiscale ships,we improve the state-of-the-art feature pyramid network(FPN)Bi FPN with channel attention module(CAM)to complete the fusion operations.Finally,we use a decoupled head structure to deal with the ship classification and bounding box regression tasks separately.Based on these novel techniques,our experiments on the public SAR datasets show that the improved model outperforms the mainstream detectors and provides a new solution for multi-scale ship target detection in complex backgrounds.In order to improve the feature fusion process and enhance the feature representation of ship objects,in the meantime,to optimize the label assignment and lift the detection performance of detectors,we introduce a novel anchor-free model with salient feature fusion mechanism for ship detection in SAR images.The network of the anchor-free algorithm YOLOX is redesigned by combining a saliency region extraction(SRE)module which is applied to generate the corresponding salient guide map of the input image;Besides,the salient feature fusion(SFF)module is proposed to obtain the fused features via deep feature maps and salient feature maps which highlight the salient regions of ship targets.Finally,we improve the Sim OTA mechanism by introducing the anchor Io Us,which will shield the adverse effect of the inaccurate predictions in the early stage of training for the tasks of object classification and bounding box regression.The experiments are conducted on public datasets which prove the model’s effectiveness for ship detection task in SAR images and high generalization ability when the detection task changes.For the purpose of enriching the diversity of features,fully extracting the context information,and reducing the interference from the neighboring targets,we propose an improved anchor-free model based on dilated attention mechanism for ship detection in SAR images.On the basic framework of the off-the-shelf YOLOX,a lightweight dilated convolutional attention module(DCAM)is embedded in front of the FPN to adjust the relationship between receptive field and multi-scale feature fusion,and strengthen the representation ability of features;The center-ness prediction branch is designed to constrain low-quality detection boxes that are far from the target center,and the branch shares the same network with the bounding box regression branch to optimize the detection results without introducing additional parameters.The comparison experiments on the public SAR datasets demonstrate that the proposed model is prone to reducing the interference from complex backgrounds and suitable for the multi-scale ship detection in SAR images.