Synthetic Aperture Radar Ship Detection Method Based on Highly Efficient Aggregated Feature Enhancement Network
Objective Synthetic aperture radar(SAR)is a microwave imaging radar that utilizes the principle of synthetic aperture to achieve high resolution.It has various characteristics such as all day,all weather,high resolution,and wide bandwidth.It is not affected by weather,day,and night and can obtain high-quality,high-resolution,large-scale,and long-distance images.SAR ship target detection technology can provide important technical support in industries such as ocean,oil,port management,marine resource development,and marine scientific research,as it can detect ships and equipments on the sea and detect potential safety risks in advance.At the same time,ship target detection technology has important strategic significance for strengthening maritime monitoring,border patrol,maritime rescue,and safety assurance of maritime channels.We aim to improve the accuracy of SAR ship detection,reduce false positives,and enhance the adaptability of the model.Methods Traditional SAR image target detection methods include texture analysis,polarization characteristics,and constant false alarm rate(CFAR)algorithms.Among them,the most widely used is the CFAR detection algorithm,which has certain advantages in speed,but its drawbacks are high computational complexity and susceptibility to complex backgrounds,resulting in unsatisfactory detection efficiency.In the actual SAR imaging process,the backgrounds of SAR images are mostly ports,islands,reefs,and other buildings.These backgrounds have high grayscale characteristics and strong confusion.Therefore,for the detection of ship targets on the sea,multiple complex backgrounds,various irregular arrangements of ships,similar target misdetection,and other uncertain factors should be considered.The target features of uncertain factors have a certain degree of similarity with ships.Therefore,we propose an efficient aggregation feature enhancement network(EAFENet)to solve the problems of low accuracy,serious false detections,and unstable effects in current SAR ship target detection.The core idea is to efficiently aggregate stacking modules and introduce residual structures to effectively transmit gradient and feature information and alleviate the problem of gradient vanishing and feature loss.The combination of the CBS(convolution+batch normalization+SiLU)module,CBAM(channel spatial attention mechanism),and leaky ReLU activation function increases the sensitivity of the network to target features and introduces low-dimensional feature fusion.Through multi-layer feature pyramid connections,the expression of features is further extended and enhanced,and the residual idea is used for skip connections,enhancing the learning ability and generalization of the model.Results and Discussions In this article,qualitative and quantitative experiments are conducted on EAFENet and other mainstream models for detecting SAR ships,as well as ablation experimental analysis.To demonstrate the effectiveness of each improvement point in this article,the YOLOv7 network model is used as a benchmark,and six sets of experiments are conducted on the SSDD dataset,with the same environment and parameters.The detected images include multiple targets,few targets,and complex backgrounds.As shown in Table 3,the effect is not ideal only when attention is used alone,and the effect is significantly improved when the mentioned EL-CB(efficient layer convolutional block)is used.The proposed global enhanced feature pyramid network branch structure is used to improve the performance of the feature pyramid and enhance the fusion of shallow features.The accuracy is improved by nearly three percentage points;the recall rate and mAP0.50:0.95 are both improved by nearly 10 percentage points,and mAP0.5 is improved by 6.4 percentage points,proving the effectiveness of each module in this article.In order to further compare the performance of the proposed model,the improved algorithm is used for comparative experiments with the current mainstream algorithms.The experimental environment is the same,and the same training and testing sets are used.The indicators of Faster R-CNN,SSD,YOLOv5,YOLOv7,CenterNet,and our algorithm are shown in Table 4.In terms of accuracy,the EAFENet model is more prominent than other mainstream algorithms.EAFENet performs the best with an accuracy of 95.40％,followed by YOLOv5 and YOLOv7,with an accuracy of 93.32％and 92.90％,respectively.The accuracy of SSD and Faster R-CNN is 84.10％and 82.70％,respectively.Compared with other algorithms,EAFENet uses a more efficient feature extraction module,which to some extent reduces misjudgment.However,mainstream algorithms such as SSD have relatively weak designs in feature extraction and other aspects,as well as a lack of deeper fusion of shallow features in the feature fusion process,resulting in relatively inaccurate prediction results.Therefore,when considering the mAP value,EAFENet still performs the best,reaching 98.90％of mAP0.5,followed by YOLOv5 and YOLOv7,reaching 94.25％and 92.50％,respectively.The mAP0.5 of SSD and Faster R-CNN is 86.01％and 89.17％,respectively.However,the proposed algorithm has undergone deeper fusion in the network structure,resulting in a slight decrease in FPS(frame per second).Overall,compared with other classic algorithms,the proposed algorithm still has significant advantages in speed,and the greatly reduced false detection rate can meet the basic needs of real-time detection.Conclusions In response to the problems of low accuracy and high false detection rate in SAR ship detection,we propose a SAR ship detection method based on an EAFENet.An EL-CB is constructed through spatial channel attention as the feature extraction module of the backbone network,and Inception NeXt is used as the feature extraction part of neck to improve algorithm efficiency,enabling the network model to better understand multi-scale information with detail perception ability.In the network structure,a global enhanced feature pyramid branch structure is constructed by fusing deep-level features with low-level features.This enables the feature extraction network to simultaneously consider both low-level and deep-level information,effectively enhancing the ability to obtain features and ensuring better stability for ship detection in complex backgrounds.The experimental results show that compared with various current detection algorithms,the proposed algorithm has higher detection accuracy and can meet the needs of real-time detection.In future research,the network structure will be further optimized to improve detection accuracy and efficiency.
NETL
NSTL
万方数据
