基于改进YOLOv7的SAR图像舰船目标检测算法

An Improved YOLOv7 Based Algorithm for Ship Target Detection in SAR Images

针对SAR舰船数据集小物体在图像中像素占比小、物体识别不清、检测效率低等问题,提出一种改进YOLOv7的SAR舰船目标检测算法STSD-YOLO。首先,根据SAR图像特点,重新设计网络结构,改变多尺度特征融合与特征提取的关系,解决下采样次数过多而丢失细节特征的问题;然后,使用轻量型注意力机制Shuffle Attention,在空间域与通道域注意力机制基础上,融合特征分组与通道置换,提升网络特征提取能力,降低计算复杂度;其次,引入卷积变体DSConv,通过在可变量化内核中仅储存整数来实现减少计算量;最后,加入NWD度量,将边界框建模为2D高斯分布,以衡量小物体的边界框之间的相似性来增强对小物体的检测性能。使用HRSID舰船数据集进行了实验验证,结果表明,相较于基准算法,所提STSD-YOLO算法在舰船检测任务中mAP提升9.9%,模型体积下降62.55%。通过对比实验验证,所提改进算法对比其余主流算法检测效果更优,能有效解决SAR图像检测的问题,可以胜任SAR图像中的舰船检测任务。

In SAR ship dataset,small objects only account for a small proportion of the image's pixels,the objects cannot be clearly recognized,and the detection efficiency is low.To solve the problems,an improved YOLOv7 based SAR ship target detection algorithm STSD-YOLO is proposed.Firstly,according to the characteristics of SAR images,the network structure is redesigned,and the relationship between multi-scale feature fusion and feature extraction is changed to solve the problem of losing detailed features due to excessive down-sampling times.Secondly,a lightweight attention mechanism named Shuffle Attention is used to fuse feature grouping with channel replacement based on the spatial domain and channel domain attention mechanism to improve the feature extraction ability of the network and reduce computational complexity.Then,the convolution variant DSConv is introduced to reduce computational intensity by storing only integers in the variable quantization kernel.Finally,the NWD metric is added to improve the performance of small object detection by modeling the bounding box as a 2D Gaussian distribution to measure the similarity between the bounding boxes of small objects.The HRSID ship dataset is adopted for experimental verification.In comparison with the baseline algorithm,the STSD-YOLO algorithm has its mAP increased by 9.9%in the ship detection task,and model volume reduced by 62.55%.Through comparative experiments,it is shown that the improved algorithm has better detection effects than other mainstream algorithms.It can effectively address the difficulties of SAR image detection,which is competent to carry out the ship detection task in SAR images.