轻量化目标检测模型SSD-MV3(Single Shot Detection-MobileNet V3)因输入图像尺寸限制无法直接检测高分辨率大尺寸合成孔径声纳(Synthetic Aperture Sonar,SAS)图像中感兴趣小目标.为此,本文提出了一种新的目标检测方法HRSSD(High Resolution Single Shot Detection),该方法通过冗余切割确保SSD-MV3输入图像尺寸的规范以及感兴趣小目标的完整,并利用二次非极大值抑制保证检测结果的唯一.此外,提出了一种尺度、空间和通道注意力机制联合的特征提取模块,并利用该模块重新设计了SSD-MV3的基础网络和附加特征提取网络,记作SSD-MV3P(Single Shot Detec-tion-MobileNet V3 Pro),使得SSD-MV3P能更有效的感知感兴趣小目标特征信息.实验结果表明,在感兴趣小目标检测数据集SST(Sonar Small Targets)上,SSD-MV3P的平均检测精度(mean Average Precision,mAP)比SSD-MV3提升4.39%.HRSSD实现了高分辨率大尺寸SAS图像感兴趣小目标的检测,并且保证了同一位置上检测结果的完整性和唯一性.
Interested Small Target Detection Method Based on Improved SSD for Synthetic Aperture Sonar Image
The efficient object detection model SSD-MV3(Single Shot Detection-MobileNet V3)cannot directly de-tect the interested small targets in high-resolution SAS(Synthetic Aperture Sonar)images due to the input image size limit.To this end,this paper proposes a novel object detection method,HRSSD(High Resolution Single Shot Detection),which ensures the specification of SSD-MV3 input image size and the integrity of the interested small targets through redundant cutting algorithm,and guarantees the unique detection result by using secondary non-maximum suppression.Furthermore,an improved feature block with a combination of scale,space and channel attention mechanism is proposed,and the basic network and additional feature network of SSD-MV3 are redesigned as SSD-MV3P(Single Shot Detection-MobileNet V3 Pro).Thus,SSD-MV3P can more effectively perceive the feature information of interested small targets.The experimental results show that the mAP(mean Average Precision)of SSD-MV3P is 4.39%higher than that of SSD-MV3 on the interest-ed small target detection dataset SST(Sonar Small Target).HRSSD realizes the detection of the interested small targets in high-resolution SAS images,and ensures the integrity and uniqueness of the detection result at the same location.
synthetic aperture sonarinterested small target detectionefficient object detection modelattention mechanismsecondary non maximum suppression
国家科技期刊平台
NSTL
万方数据
