计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (11): 150-159.DOI: 10.3778/j.issn.1002-8331.2010-0171

• 模式识别与人工智能 • 上一篇    下一篇

融入特征融合与特征增强的SSD目标检测

刘建政,梁鸿,崔学荣,钟敏,李传秀   

  1. 中国石油大学(华东) 计算机科学与技术学院,山东 青岛 266580
  • 出版日期:2022-06-01 发布日期:2022-06-01

SSD Visual Target Detector Based on Feature Integration and Feature Enhancement

LIU Jianzheng, LIANG Hong, CUI Xuerong, ZHONG Min, LI Chuanxiu   

  1. College of Computer Science & Technology, China University of Petroleum(East China), Qingdao, Shandong 266580, China
  • Online:2022-06-01 Published:2022-06-01

摘要: 针对SSD算法在目标检测过程中对小目标检测的不足,提出了一种基于SSD算法的一阶段目标检测器——FIENet(feature integration and feature enhancement network)。在FIENet中设计了两个模块,一是特征融合模块,该模块对SSD浅层的特征映射信息进行融合以提高小目标检测能力;二是特征增强模块,该模块采用了残差网络(Res2Net)以及注意力机制(attention),对特征融合后的模块以及SSD中的深层特征映射进行增强。为了更好地检测小目标,还调整了浅层特征映射先验框的数量。为了评价FIENet的有效性,在PASCAL VOC2007以及MSCOCO数据集上进行了实验。实验结果表明,在PASCAL VOC2007数据集上检测精度(mAP)较SSD提高3.1个百分点,对小目标bird、bottle、chair、plant检测精度分别提升了3.6、9.5、5.4、5.5个百分点。在COCO数据集上达到29.4%的检测精度(mAP)。实验结果证明FIENet网络在保持实时性的同时可以达到较高的检测精度。

关键词: 小目标, 特征融合, SSD, 特征增强

Abstract: A one-stage target detector based on SSD algorithm is proposed, named the FIENet(feature integration and feature enhancement network), to solve the deficiency of said SSD algorithm in detecting small targets. Two blocks are designed in FIENet. One is a feature integration block, which fuses the feature mapping information in the shallow layer of SSD to improve the detection ability of small targets. The other is a feature enhancement block, which adopts the residual network(Res2Net) and the attention mechanism to enhance the feature integration block and the deep feature mapping in SSD. In order to better detect small targets, the number of prior boxes for shallow feature mapping is adjusted. In order to evaluate the effectiveness of FIENet, experiments are carried out on the PASCAL VOC2007 and MS COCO data sets. The experimental results show that the mean average precision(mAP) of the PASCAL VOC2007 data is 3.1?percentage points higher than that of SSD, and the detection accuracy of birds, bottles, chairs, and plants is 3.6, 9.5, 5.4, and 5.5?percentage points higher than that of SSD. Meanwhile, a mAP of 29.4% is achieved on the COCO data set. These results indicate that the FIENet can achieve high detection accuracy while maintaining real-time performance.

Key words: small target, feature integration, SSD, feature enhancement