面向目标检测的SAR图像去噪和语义增强  

SAR Image Denoising and Semantic Enhancement for Object Detection

作　　者：曲海成[1] 申磊 QU Haicheng;SHEN Lei(College of Software,Liaoning Technical University,Huludao,Liaoning 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《光子学报》2022年第4期321-335,共15页Acta Photonica Sinica

基　　金：国家自然科学基金(No.42071351);辽宁省教育厅基础研究项目(No.LJ2019JL010);辽宁工程技术大学学科创新团队资助项目(No.LNTU20TD-23)。

摘　　要：为了解决合成孔径雷达图像中舰船目标容易在复杂背景下被淹没、相干斑噪声导致舰船边缘模糊和小尺度舰船目标在经过多次卷积后容易丢失的问题,提出像素级去噪和语义增强的检测模型。首先,利用预测掩码逐像素指导特征图,激活目标信息,抑制背景和相干斑噪声,检测相干斑噪声和相似舰船目标影响的图像。其次,利用语义增强模块增强特征图中语义信息,使得每层特征图都包含丰富的语义信息,进而判断小尺度舰船。最后,引入Transformer Encoder模块,提高舰船目标和特征图之间的上下文信息,增强舰船目标和图像之间的依赖关系。提出的模型能够有效减少漏检、误检情况,在公开数据集SSDD上进行测试,检测精度达到96.73%,其中针对小尺度舰船检测精度达到96.85%,大尺度舰船检测精度达到96.41%,远海场景检测精度达到98.53%,近海岸场景检测精度达到90.00%,验证了该模型的有效性和泛化能力。The imaging process of the synthetic aperture radar system is not affected by time and weather,and can achieve all-day and all-weather imaging of the target. It has a wide detection range and generates high-resolution images. Therefore,it is widely used in the military and civilian fields. In recent years,satellites of "GF-3" and satellites of "HJ-1" were successively launched to fill the gap in synthetic aperture radar technology in China. However,synthetic aperture radar images still have shortcomings in the target detection process. Ship targets in SAR images are sparse and of various scales. The anchor box-based detection model relies too much on manually designed candidate boxes,which cannot adapt to all ship targets,and the parameters of the candidate boxes consume a lot of computing resources. The background of the synthetic aperture radar image is complex,and the ship target can easily disappear in the complex background,which leads to the missed detection of the detection model. SAR images contain a large number of small-scale ship targets,which are easily lost after multiple convolutions. Coherent speckle noise present in SAR images causes blurring of ship edges. To address the impact of the above problems,this paper proposes a detection model for pixel-level denoising and semantic enhancement. First,the pixellevel denoising module uses the prediction mask to generate the attention map of[0,1],multiplies the feature map and the attention map pixel by pixel to achieve denoising,and optimizes the attention map using the cross-entropy loss. The denoising module can enhance the weight of the target area of the ship,suppress the weight of the non-target area,and enhance the difference between the ship target information and the background information in the feature map. Second,the semantic enhancement module enhances the semantic information contained in the feature map,and uses asymmetric convolutional layers to extract features of different dimensions,preventing candidate boxes with high IOU scores and low

关 键 词：目标检测 图像处理 深度学习 SAR图像 像素级去噪 语义增强 TRANSFORMER 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]