计算机与现代化 ›› 2021, Vol. 0 ›› Issue (06): 18-23.

• 算法设计与分析 • 上一篇    下一篇

DF-SSD:一种基于反卷积和特征融合的单阶段小目标检测算法

  

  1. (1.贵州省先进计算与医疗信息服务工程实验室(贵州大学计算机科学与技术学院),贵州贵阳550025;
    2.航天江南集团有限公司,贵州贵阳550009; 3.贵州联科卫信科技有限公司,贵州贵阳550001)
  • 出版日期:2021-07-05 发布日期:2021-07-05
  • 作者简介:王良玮(1995—),女,贵州遵义人,硕士研究生,研究方向:大规模数据处理与分析,计算机视觉,E-mail: 2442187702@qq.com; 陈梅(1964—),女,教授,博士,研究方向:大规模数据管理与分析,医学信息管理与分析服务; 通信作者:李晖(1982—),男,教授,博士,研究方向:大规模数据管理与分析,人工智能技术及应用,高性能数据库,云服务,E-mail: cse.HuiLi@gzu.edu.cn; 李焕军(1980—),男,高级工程师,研究方向:工业大数据与工业人工智能; 施若(1966—),男,博士,研究方向:多源医疗数据融合处理; 戴震宇(1985—),男,实验师,研究方向:大规模数据管理与分析。
  • 基金资助:
    国家自然科学基金资助项目(71964009); 贵州省高层次创新型人才项目(黔财教[2018]190)

DF-SSD: A One-stage Small Target Detection Algorithm Based on Deconvolution and Feature Fusion

  1. (1. Guizhou Engineering Laboratory for Advance Computing and Medical Information Service College of (Computer Science and 
    Technology, Guizhou University), Guiyang 550025, China; 2. Aerospace Jiangnan Group Co. Ltd., Guiyang 550009, China; 
    3. Guizhou Lianke Weixin Technology Co. Ltd., Guiyang 550001, China)
  • Online:2021-07-05 Published:2021-07-05

摘要: 针对经典的单阶段多目标检测算法SSD对小目标物检测效果差的问题,提出DF-SSD算法,其核心技术贡献包括基于反卷积与特征融合的方法和改进后的先验框尺寸计算算法。反卷积与特征融合能够增加浅层特征层的语义信息。改进后的先验框尺寸计算引入了数据集的特点,能有效利用每一个先验框进行训练和预测。改进后的方法DF-SSD与基于SSD改进的R-SSD和RSSD模型相比,具有较高的检测准确率。同时,DF-SSD的检测时间仅是R-SSD的1/2,是DSSD的1/5。改进后的方法在VOC2007和DIOR这2个数据集上的MAP比SSD分别提升了1.4和3.6个百分点。其中ship、vehicle、windmill、cat这4类小目标的MAP分别提升了23.2、12.6、8和4.8个百分点。结果表明:DF-SSD方法有效提高了小目标物的检测正确率,并且具有较快的检测速度。

关键词: SSD模型, 反卷积, 特征融合, 小目标检测, PASCAL VOC2007, DIOR

Abstract: Aiming at the problem of the SSD model’s poor detection performance on small targets, the DF-SSD algorithm was proposed, its technical contributions include a one-stage detector method based on deconvolution and feature fusion and an improved default bounding boxes’ size calculation algorithm. Deconvolution and feature fusion can increase the semantic information of shallow feature layers. In DF-SSD algorithm, the improved default bounding boxes’ size calculation introduces the characteristics of the data set, which can effectively use each default bounding box for training and prediction. Compared with the improved R-SSD and RSSD models based on SSD, the DF-SSD method has higher detection accuracy. At the same time, DF-SSD’s detection overhead is only 1/2 of R-SSD and 1/5 of DSSD. The MAP of the DF-SSD on the VOC2007 and DIOR data sets is 1.4 and 3.6 percentage points higher than that of SSD respectively. Meanwhile, DF-SSD’s MAP of small targets of ship, vehicle, windmill, and cat increased 23.2, 12.6, 8 and 4.8 percentage points respectively. The results show DF-SSD effectively improves the detection accuracy of small targets and has a faster detection speed.

Key words: SSD model, deconvolution, feature fusion, small target detection, PASCAL VOC2007, DIOR