计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (17): 158-164.DOI: 10.3778/j.issn.1002-8331.2111-0441

• 模式识别与人工智能 • 上一篇    下一篇

全卷积目标检测的改进算法

廖永为,张桂鹏,杨振国,刘文印   

  1. 1.广东工业大学 计算机学院,广州 510006
    2.鹏城实验室 网络空间安全研究中心,广东 深圳 518000
  • 出版日期:2022-09-01 发布日期:2022-09-01

Improved Algorithm for Fully Convolutional Object Detection

LIAO Yongwei, ZHANG Guipeng, YANG Zhenguo, LIU Wenyin   

  1. 1.School of Computer, Guangdong University of Technology, Guangzhou 510006, China
    2.Cyberspace Security Research Center, Peng Cheng Laboratory, Shenzhen, Guangdong 518000, China
  • Online:2022-09-01 Published:2022-09-01

摘要: 基于无锚点的单阶段全卷积目标检测算法(FCOS)无需生成大量的锚点避免了样本不平衡问题,但FCOS可能更适应于某一特定场景。为了增强特征融合,并提高目标检测的准确性,提出了全卷积目标检测算法FCOS的改进算法ConFCOS。该算法设计了一个增强的特征金字塔网络,引入带有全局上下文信息的注意力模块和空洞卷积模块,以减少特征融合过程中的信息衰减。另外,构建了一个级联检测头来检测对象,对检测的边界框进行细化来提高分类和回归的置信度。此外,针对提出的ConFCOS的损失函数进行了优化以提高目标检测的准确率。在COCO数据集上进行的实验表明,ConFCOS的准确度比FCOS提高了1.6个百分点。

关键词: ConFCOS, 增强的特征金字塔网络, 级联检测, 目标检测

Abstract: The anchor-free fully convolutional one-stage objectdetection(FCOS) avoids the problem of sample imbalance because it does not generate a large number of anchors. However, FCOS may be applicable to specific scenario. To improve feature fusion and accuracy of object detection, an improved algorithm for fully convolutional object detection named ConFCOS is proposed, which designs a strengthened feature pyramid network(SFPN) with the attention modules of global context information and dilated convolutional modules to reduce information attenuation in feature fusion. Moreover, a cascade detection head is constructed to detect the objects, which uses the cascade detection to refine the bounding box regression and improve the confidence of classification and regression. In addition, the loss function of ConFCOS is optimized to accelerate model training and improve performance of detector. Experimental results on COCO datasets show that the mAP of ConFCOS is about 1.6?percentage points higher than FCOS.

Key words: ConFCOS, strengthened feature pyramid network(SFPN), cascade detection, object detection