摘要
小目标检测的主要任务是检测图像中尺寸小于32×32像素的目标并对其分类。由于传统矩形锚框结构检测小目标时匹配不准确,小目标在通用数据集中数量较少且分布不均匀,导致模型检测效果较差。为此,在Faster RCNN的基础上,提出一种圆形锚框的小目标检测方法。在RPN阶段采用圆形锚框定位感兴趣区域,通过新的面积交并比计算方法与损失函数减少模型参数量与锚框回归阶段的偏移计算,以增强模型对被检测目标的拟合能力,提升模型检测精度和效率。同时,为了解决现有公开数据集中小目标占比较少及分布不均匀问题,在MS COCO 2017数据集上进行数据增强操作,仅保留其中的小目标并将标注信息修改为对小目标包裹率较高的圆形包围框。实验表明,采用圆形锚框方法与数据增强方法在检测小目标时检测效果较好,检测效率、速度均明显优于Faster R-CNN,APS、检测速度分别提升4.1%与4 FPS。
The main task of small object detection is to detect images with dimensions smaller than 32×32 pixel target and classify it.Due to the inaccurate matching of traditional rectangular anchor frame structures in detecting small targets,the number of small targets in the general dataset is small and their distribution is uneven,which will lead to poor model detection performance.Therefore,based on Faster R-CNN,a small target detection method with circular anchor frames is proposed.In the RPN stage,a circular anchor frame is used to locate the region of interest,and a new area intersection and union ratio calculation method and loss function are used to reduce the model parameter quantity and offset calculation in the anchor frame regression stage,in order to enhance the model's fitting ability to the detected target and improve the model's detection accuracy and efficiency.At the same time,in order to address the issues of low proportion and uneven distribution of small targets in existing public datasets,data augmentation was performed on the MS COCO 2017 dataset,retaining only the small targets and modi-fying the annotation information to a circular bounding box with a high wrapping rate for the small targets.Experiments have shown that the cir-cular anchor box method and data augmentation method have better detection performance in detecting small targets,with detection efficiency and speed significantly better than Faster R-CNN,APS and detection speed have been improved by 4.1%and 4 FPS,respectively.
作者
闫春相
徐遵义
刘康宁
李晨
YAN Chunxiang;XU Zunyi;LIU Kangning;LI Chen(College of Computer Science and Technology,Shandong Jianzhu University,Ji'nan 250101,China)
出处
《软件导刊》
2024年第1期128-134,共7页
Software Guide
基金
国家自然科学基金青年基金项目(62102235)
山东省重点研发计划(重大科技创新工程)项目(2021CXGC011204)。
